2026-10-21T1148Z-audit
One pipeline fire, in full · audit run of 2026-10-21 · sub-agent allocation and telemetry, per-iteration verification verdicts and findings, source-list edits, coverage gaps, bridge invocations, and the run's own verification & coverage notes: what was published, what was dropped at the borderline or judged not relevant (and why), single-source carve-outs, and contradictions. Rendered from runs/2026-10-21/2026-10-21T1148Z-audit.md.
Run telemetry
Claude Opus 5.5
(claude-opus-5-5)
main agent
repair-1
Claude Opus 5.5
(claude-opus-5-5, container default, env fallback)
- Items returned
- 0
- Duration
- 290m 41s
- Tool calls
- not reported
- Cited sources
- none
repair-2
Claude Opus 5.5
(claude-opus-5-5, container default, env fallback)
- Items returned
- 0
- Duration
- 288m 48s
- Tool calls
- not reported
- Cited sources
- none
repair-3
Claude Opus 5.5
(claude-opus-5-5, container default, env fallback)
- Items returned
- 0
- Duration
- 289m 51s
- Tool calls
- not reported
- Cited sources
- none
repair-4
Claude Opus 5.5
(claude-opus-5-5, container default, env fallback)
- Items returned
- 0
- Duration
- 291m 52s
- Tool calls
- not reported
- Cited sources
- none
truth-1
Claude Sonnet 5.5
(claude-sonnet-5-5 until the session limit, claude-opus-5-5 after the resume (self-reported))
- Items returned
- 0
- Duration
- 246m 08s
- Tool calls
- not reported
- Cited sources
- none
truth-2
Claude Sonnet 5.5
(claude-sonnet-5-5 until the session limit, claude-opus-5-5 after the resume (self-reported))
- Items returned
- 0
- Duration
- 229m 22s
- Tool calls
- not reported
- Cited sources
- none
truth-3
Claude Sonnet 5.5
(claude-sonnet-5-5 until the session limit, claude-opus-5-5 after the resume (self-reported))
- Items returned
- 0
- Duration
- 237m 16s
- Tool calls
- not reported
- Cited sources
- none
truth-4
Claude Sonnet 5.5
(claude-sonnet-5-5 until the session limit, claude-opus-5-5 after the resume (self-reported))
- Items returned
- 0
- Duration
- 232m 02s
- Tool calls
- not reported
- Cited sources
- none
Verification
#? NEEDS_FIXES · Sonnet 5.5 · t=62 e=29 a=26 #? NEEDS_FIXES · Sonnet 5.5 · t=9 e=3 a=4
Deep dive
·
Entries this run published (0) and updated (74)
- Groupe 3R (Réseau Radiologique Romand): Akira ransomware attack on 20 Swiss imaging centres, administrative data stolen and published, medical data encrypted
- Braintrust AI evaluation platform AWS account breach: customer-stored AI provider API keys likely exposed, all customers urged to rotate org-level keys
- JDownloader website compromise: "Download Alternative Installer" and Linux shell installer links pointed to malware on 6 and 7 May 2026
- Bauman University "Department No. 4": leaked documents show a GRU training pipeline that places graduates in Sandworm and Fancy Bear units
- macOS infostealer campaign shifts to ClickFix Terminal commands to deliver MacSync, SHub Stealer and AMOS
- cPanel & WHM security update of 2026-05-08: CVE-2026-29202 Perl code injection, CVE-2026-29203 symlink chmod and CVE-2026-29201 arbitrary file read
- BSI lists two authenticated pfSense CE code-execution flaws as critical: CVE-2025-69690 (CE 2.7.2) and CVE-2025-69691 (CE 2.8.0), disputed by Netgate
- SMS-blaster smishing in Switzerland: fake base stations push phones onto 2G and send spoofed texts past operator filters
- ICO fines South Staffordshire Water £963,900 under UK GDPR for the 2020 to 2022 intrusion: phishing foothold, ZeroLogon on two unpatched domain controllers, 5% of IT monitored
- BKA and ZIT shut down the relaunched Crimenetwork darknet marketplace; suspected German operator arrested in Mallorca on a European Arrest Warrant
- West Pharmaceutical Services 8-K Item 1.05: data exfiltrated and systems encrypted, global operations partly restarted
- Škoda Auto Deutschland accessories-shop breach: attackers exploited the shop software, customer data and password hashes were accessible, logs cannot show what was copied
- Checkmarx Jenkins AST plugin backdoored: rogue build 2026.5.09 swept build hosts for credentials on 2026-05-09/10, the third wave of malicious Checkmarx artifacts since March (TeamPCP claimed)
- GTIG AI Threat Tracker (May 2026): an AI-developed zero-day 2FA bypass prepared for mass exploitation, AI decoy code in Russia-nexus malware and PROMPTSPY's Gemini agent
- Foxconn confirms a cyberattack on some North American factories after Nitrogen claims theft of 8 TB of customer project data
- BWH Hotels (Best Western, WorldHotels, Sure Hotels): attackers had access to a guest-reservation web application for about six months
- CVE-2026-44277 and CVE-2026-26083: unauthenticated code execution in the Fortinet FortiAuthenticator API and the FortiSandbox web UI
- CVE-2026-45185 "Dead.Letter": Exim use-after-free in BDAT/CHUNKING on GnuTLS builds, fixed in 4.99.3
- CVE-2026-34263 and CVE-2026-34260: SAP Commerce Cloud unauthenticated code injection and S/4HANA Enterprise Search SQL injection (May 2026 Patch Day)
- CERTFR-2026-AVI-0564: SPIP before 4.4.14, two remote code execution flaws in the private area and, on certain nginx configurations, the public area
- CERTFR-2026-AVI-0572: Centreon April 2026 bulletin, SQL injection in eight components, command injection in MBI and XSS in MAP
- Microsoft MDASH: multi-model agentic scanning harness credited with 16 Windows networking and authentication CVEs, four of them Critical remote code execution flaws
- TrickMo C: Android banking trojan moves its C2 onto The Open Network and turns infected phones into SOCKS5 and SSH network pivots
- NCSC-UK: 10 questions to ask when using AI models to find vulnerabilities
- Mini Shai-Hulud hits TanStack: a pull_request_target Pwn Request, Actions cache poisoning and OIDC token theft from runner memory published 84 malicious @tanstack versions with valid provenance
- Dutch IGJ: Clinical Diagnostics labs failed NEN 7510 at the time of the July 2025 Nova breach, which reached data of about 941,000 people through an account without MFA in unmonitored legacy systems
- CVE-2026-8043, Ivanti Xtraction: authenticated file read and HTML write (CVSS 9.6), patched with an Endpoint Manager SQL injection to RCE and a Virtual Traffic Manager admin command injection in Ivanti's May 2026 batch
- GemStuffer: researchers attribute the May 2026 RubyGems spam campaign to OpenAI agents that ran code on RubyDoc.info and tried to steal API keys; RubyGems cannot confirm and OpenAI disputes it
- The Gentlemen RaaS: leaked internal chats show Fortinet and Cisco edge access, NTLM relay, ADCS abuse and EDR-kill tooling, mapped by Check Point Research
- FamousSparrow's three-wave intrusion at an Azerbaijani oil and gas company: repeated ProxyNotShell access, Deed RAT behind a two-export LogMeIn Hamachi sideload, and a blocked TernDoor attempt
- FrostyNeighbor (Ghostwriter, UNC1151): since March 2026, geofenced PDF lures deliver a JavaScript PicassoLoader and Cobalt Strike to Ukrainian government organisations
- CVE-2026-45691, Nextcloud Server and Enterprise Server: second-factor bypass on DAV endpoints by reusing a pre-2FA session cookie (CVSS 5.9)
- CVE-2026-45793, PHP Composer: GitHub Actions and GitHub App installation tokens printed in full to CI job logs when GitHub's new token format failed validation (CVSS 7.5)
- Sophos State of Identity Security 2026: Swiss respondents report the highest identity-breach rate of 17 countries, at 89%
- Datadog analyses TeamPCP's open-sourced Shai-Hulud framework: credential and runner-memory theft, IDE and Claude Code hook persistence, OIDC npm publishing and a wipe-on-revocation daemon
- Suspected Dream Market main administrator "Speedstepper" arrested in Germany; US indictment traces reactivated marketplace wallets and gold purchases
- OpenClaw "Claw Chain": four chainable sandbox, exec-allowlist and MCP loopback flaws (CVE-2026-44112, CVE-2026-44113, CVE-2026-44115, CVE-2026-44118) fixed in release 2026.4.22
- AMD-SB-7052 / CVE-2025-54518: AMD Zen 2 op-cache corruption lets local code run instructions at a higher privilege, guest to host included (CVSS 7.3); OS updates on EPYC 7002, OEM BIOS updates on Ryzen
- Unit 42: Gremlin stealer hides its payload in XOR-encoded .NET resources and adds a crypto clipper, Discord token theft and WebSocket session hijacking
- SentinelOne "Living Off the Pipeline": attackers subverted TeamCity, GitLab automation, self-hosted runners and developer workstations in 2025
- CVE-2026-44088: SzafirHost, KIR's qualified e-signature helper, verifies a JAR's signature and loads its classes from different parts of the file (fixed in 1.2.1)
- FunnelKit Funnel Builder for WooCommerce: unauthenticated settings write (CVE-2026-47100) exploited to plant card skimmers on checkout pages, fixed in version 3.15.0.3
- CVE-2026-41225: F5 BIG-IP iControl REST lets an authenticated Manager-role account run arbitrary commands and bypass Appliance mode (May 2026 quarterly)
- CVE-2026-41553: DHTMLX PDF Export Module executes JavaScript injected into its data parameter without authentication (CVSS 4.0 10.0), with two file-read flaws (CVE-2026-41552, CVE-2026-7182)
- Kaspersky: Kimsuky adds Rust-based HelloDoor and httpMalice to its PebbleDash toolkit and takes over hosts through VSCode tunnels and DWAgent
- Tycoon 2FA after the March 2026 takedown: operators add OAuth device-code phishing against Microsoft 365 alongside the credential-relay kit
- ARWINI, Lower Saxony's prescription audit body: police confirm a data outflow and name the Kairos group, which lists 2.87 TB for sale
- BigBlueButton bbb-web before 3.0.21 and 3.0.23: predictable session tokens, an API checksum bypass and an SSRF (CVE-2026-46351, CVE-2026-46353, CVE-2026-46404)
- CISA contractor's public GitHub repository exposed AWS GovCloud admin keys and plaintext passwords for internal CISA systems for up to six months
- 7-Eleven confirms breach of franchisee documents after ShinyHunters claims 600,000+ Salesforce records
- INTERPOL Operation Ramz: 13 MENA countries arrest 201 suspects, seize 53 servers and dismantle a phishing-as-a-service site in Algeria
- Symantec and Carbon Black: the fast16 sabotage framework corrupted LS-DYNA and AUTODYN nuclear-detonation simulations; Kim Zetter argues it was contemporaneous with Stuxnet
- Grafana Labs confirms GitHub breach claimed by CoinbaseCartel: a workflow token missed after the TanStack npm compromise let attackers download its code and internal repositories
- MiniPlasma: Chaotic Eclipse PoC turns the 2020 Windows Cloud Files driver flaw CVE-2020-17103 into a SYSTEM shell on patched Windows 11; June 2026 updates fix it
- CVE-2026-9082: anonymous SQL injection in Drupal core on PostgreSQL, Highly critical, exploited and CISA KEV-listed
- Sparx Pro Cloud Server and Enterprise Architect: five CVEs including unauthenticated SQL execution and a WebEA race to code execution, public PoC, no vendor fix at disclosure
- Nx Console VS Code extension version 18.95.0 shipped a credential stealer, published with a contributor token stolen in the TanStack npm compromise
- Previously unknown flaw in Huawei enterprise router software caused POST Luxembourg's nationwide telecom outage of July 2025; no CVE ten months later
- CVE-2026-41091: exploited Microsoft Defender link-following privilege escalation to SYSTEM, with the CVE-2026-45498 Defender denial of service
- CVE-2026-45584: Microsoft Defender heap buffer overflow gives code execution when Defender scans a malicious quarantined file
- DirtyDecrypt (DirtyCBC): public root exploit for a Linux kernel RxGK page-cache write, mapped to CVE-2026-31635, on kernels built with CONFIG_RXGK
- vm2 Node.js sandbox: two waves of critical sandbox escapes in May 2026, fixed in 3.11.2 and then 3.11.4
- Cisco Talos: the "demo.pdb" BadIIS variant is commodity IIS-module malware for SEO fraud, likely sold among Chinese-speaking crime groups and built by the author "lwxat"
- The Gentlemen ransomware group lists Czech university VSFS and Swiss engineering firm DEVO-Tech AG on its leak site
- Storm-2949: SSPR abuse of four privileged users turned into a data-theft sweep across Microsoft 365, Key Vault, SQL, Storage and Azure VMs
- Webworm (China-aligned) turns to European government targets with the EchoCreep (Discord) and GraphWorm (Microsoft Graph and OneDrive) backdoors, ESET reports
- SonicWall Gen6 SSL-VPN: patched-firmware appliances still open to the CVE-2024-12802 MFA bypass without manual LDAP remediation, exploited February to March 2026 by a suspected access broker
- B1ack's Stash carding marketplace releases about 4.6 million stolen payment card records for free
- CVE-2026-42822: improper authentication in Azure Local Disconnected Operations allows privilege escalation over the network (CVSS 10.0), fixed in ALDO 2604
- CVE-2026-45829: ChromaDB's Python FastAPI server runs an attacker's model code before checking authentication (CVSS 4.0 10.0), no fix in 1.5.9
- Keycloak 26.6.2 fixes 16 security issues, among them an OIDC login session fixation (CVE-2026-7507) and a WebAuthn execute-actions token replay (CVE-2026-37982)
- PinTheft: Linux kernel RDS zerocopy double-free (CVE-2026-43494) turned into a root exploit through io_uring fixed buffers, public PoC; V12 found RDS loaded by default only on Arch Linux among the distributions it tested
- Verizon 2026 DBIR: vulnerability exploitation overtakes stolen credentials as the leading initial access vector for the first time in 19 years
- Showboat Linux post-exploitation framework and JFMBackdoor for Windows used against telecom providers; PwC attributes the operation to Red Lamassu (Calypso)
Sources changed (this run)
Edits this run made to sources/sources.json · promotions, demotions, new candidates, and fetch-method / category / reliability / url corrections (the run record's sources_changed[]). Paginated; 10 per page.
No source-list edits recorded for this run.
Coverage gaps (this run)
Sources this run's brief needed that returned no usable content via any documented recipe. Bridge-recovered or quiet-day sources do NOT appear here. (Distinct from the independent source-accessibility probe at the foot of this section, which probes all active sources regardless of what any run needed.)
No coverage gaps in this run · every source the brief needed returned usable content via its documented recipe.
Verification findings · all iterations
Per-iteration finding detail. Each table is one verifier pass · what was flagged, how the main agent remediated it, and the outcome. Walking the tables top-to-bottom shows the verifier's debugging trail across iterations.
Iteration #? NEEDS_FIXES · 117 findings (truth=62, editorial=29, advisory=26) · Claude Sonnet 5.5 · 31m 13s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F2 generic-url | · | 2026-05-13/mini-shai-hulud-s-github-actions-pwn-request-oidc-token-thef: (low confidence) sources[] cites the whole KEV catalog feed JSON, a listing and not a per-CVE page; cite the per-CVE vendor or | s1#1 s1#1 D: CISA publishes no per-CVE KEV page; the catalog feed is the store's established KEV citation. | s2#26 D: KEV catalog feed (same reason as s1#1). | |
| F3 claim-not-supported | · | 2026-05-10/braintrust-ai-evaluation-platform-aws-account-breach-multi-t: Claim ledger 778fa5c9b3. The cited TechCrunch article says the keys are 'used by customers for accessing cloud-based AI models' | s1#2 s1#2 A: "org-level" now cited to SecurityWeek, which quotes Braintrust's notice; TechCrunch keeps the clause it carries. | |
| F3 claim-not-supported | · | 2026-05-10/jdownloader-official-site-compromised-windows-and-linux-inst: Claim ledger 60e19b45a0. The cited BleepingComputer article says the developers shared an archive of the malicious installers a | s1#3 s1#3, s1#15 A: the hash list is cited to JDownloader's own notice (its "known SHA256 checksums of the malicious substitute installer files"), and the "rather than a manifest of its own" negative is gone from the… | |
| F3 claim-not-supported | · | 2026-05-10/bauman-university-department-no-4-leaked-gru-cyber-operator: Claim ledger b365df62d2 and record summary d04d48a00c. Cited Meduza says journalists reviewed more than two thousand internal do | s1#4 s1#4 A: "the documents themselves were not published" dropped from the section and the record summary. | |
| F3 claim-not-supported | · | 2026-05-10/cpanel-whm-second-emergency-tsr-in-10-days-embargo-lifted-on: Claim ledger fc1fa4ba2a, 9eb8e57dbd, 9204f89f55, eb59eef7b6 and record summary d7b4155c21; also the body text 'cPanel's own sco | s1#5 s1#5 A: the scores are attributed to the CVE records (HackerOne as numbering authority, per the EUVD API's assigner field), not to cPanel, in the body, the section, the sourcing note and the record summary. | |
| F3 claim-not-supported | · | 2026-05-12/gtig-ai-threat-tracker-may-2026-first-confirmed-ai-generated: Claim ledger b8f25bc644. GTIG says SANDCLOCK extracts 'cloud secrets, such as AWS keys and GitHub tokens'; AI API secrets appea | s1#6 s1#6 A: "AI API secrets" dropped from the Exposure line (GTIG lists AWS keys and GitHub tokens). | |
| F3 claim-not-supported | · | 2026-05-12/teampcp-unc6780-pcpjack-ecosystem-backdoors-the-checkmarx-je: (low confidence) Claim ledger 64a62577df. Checkmarx says 'Targeted applications include crypto wallets, VPNs, AWS, and Github' | s1#7 s1#7 A: Checkmarx's list is given as "among the applications the payload targets", without the exclusion. | |
| F3 claim-not-supported | · | 2026-05-13/cve-2026-45185-exim-dead-letter-use-after-free-in-bdat-chunk: Claim ledger 6bf045433c (body), c06848ee34 (Correction section), ea02577cec (record summary). XBOW's post says the human 'didn' | s1#8 s1#8 A: body, section and record summary now say no full exploit against the PIE build was reported: the researcher's side reached a leaked stack address, the LLM never reached a leak (XBOW). | |
| F3 claim-not-supported | · | 2026-05-13/foxconn-confirms-nitrogen-ransomware-crippled-north-american: Claim ledger 4222ae1cec. The citation is 9to5Mac, which carries only the Mount Pleasant (Cybersec Guru) and Houston (AppleInsid | s1#9 s1#9 A: the factory-location list is cited to The Record, which carries it, with Foxconn's refusal to say how many were impacted. | |
| F3 claim-not-supported | · | 2026-05-13/trickmo-trickmo-c-android-banking-trojan-migrates-c2-to-the: (low confidence) Claim ledger aa78ea6ce2. ThreatFabric says takedowns are 'largely ineffective' (reduced effect), not that they | s1#10 s1#10 A: the takeaway attributes to ThreatFabric only its finding (takedowns largely ineffective, blocking less effective); the device controls stand as derived advice without a citation. | |
| F3 claim-not-supported | · | 2026-05-13/mini-shai-hulud-s-github-actions-pwn-request-oidc-token-thef: Claim ledger 5f00cdca91. The cited 2021 GitHub Security Lab post warns against checking out untrusted PR code under pull_reques | s1#11 s1#11 A: the SHA-pinning clause replaced by what GitHub Security Lab says (pull_request_target plus an explicit checkout of an untrusted PR may lead to repository compromise). | |
| F3 claim-not-supported | · | 2026-05-13/mini-shai-hulud-s-github-actions-pwn-request-oidc-token-thef: Check 2e: the two reference-page source records carry the fetch date (2026-10-21), not a publication date; these evergreen docs | s1#12 s1#12 A: the actions/cache README and npm docs records carry date null instead of the fetch date. | |
| F3 claim-not-supported | · | 2026-05-13/cve-2026-44277-cve-2026-26083-fortinet-fortiauthenticator-an: (low confidence) The two EUVD records carry dates that are the records' dateUpdated, whereas EUVD gives datePublished May 12, 2 | s1#13 s1#13 A: both EUVD source dates set to the records' publication date, 2026-05-12. | |
| F4 hallucinated-fact | · | 2026-05-10/braintrust-ai-evaluation-platform-aws-account-breach-multi-t: (low confidence) Sources say only 'suspicious spikes in AI provider usage'; none states the spikes came from attacker use of th | s1#14 s1#14 A: T1496.004 replaced by T1528 (Steal Application Access Token), matching the sourced "likely gave the attackers access to the API keys"; "abuse signal" became "signal". Braintrust's own notice… | |
| F4 hallucinated-fact | · | 2026-05-10/jdownloader-official-site-compromised-windows-and-linux-inst: (low confidence) The vendor notice carries a SHA256 table of the malicious files; no source states the vendor published no mani | s1#15 s1#3, s1#15 A: the hash list is cited to JDownloader's own notice (its "known SHA256 checksums of the malicious substitute installer files"), and the "rather than a manifest of its own" negative is gone from the… | |
| F4 hallucinated-fact | · | 2026-05-12/gtig-ai-threat-tracker-may-2026-first-confirmed-ai-generated: (low confidence) GTIG: 'our proactive counter discovery may have prevented its use'; 'likely stopped' is stronger than 'may hav | s1#16 s1#16 A: headline says its discovery "may have prevented" mass exploitation, GTIG's own hedge. | |
| F4 hallucinated-fact | · | 2026-05-12/teampcp-unc6780-pcpjack-ecosystem-backdoors-the-checkmarx-je: (low confidence) Checkmarx: 'The malicious window was 2026-05-09 01:25 UTC to 2026-05-10 08:47 UTC', the availability of the ma | s1#17 s1#17 A: headline says the credential-hunting build was live for about 31 hours (Checkmarx's malicious window), not that it hunted for 31 hours. | |
| F4 hallucinated-fact | · | 2026-05-13/cve-2026-44277-cve-2026-26083-fortinet-fortiauthenticator-an: (low confidence, pre-existing) The primary source, Fortinet PSIRT FG-IR-26-128 / FG-IR-26-136, was published 2026-05-12; event_ | s1#18 s1#18 A: event_date 2026-05-13 to 2026-05-12, Fortinet's publication date. | |
| F14 ? | · | 2026-05-12/bka-and-zit-dismantle-relaunched-crimenetwork-darknet-market: (low confidence) Claim ledger e9842622a4. The cited Help Net Security article also quotes Meywirth: 'another administrator will | s1#19 s1#19 A: "The only forward-looking statement" dropped (Help Net also quotes the BKA on another administrator). | |
| F14 ? | · | 2026-05-13/bwh-hotels-best-western-worldhotels-sure-hotels-181-day-unau: Claim ledger beebca41bb and record summary a565e60bd0. The Register: 'owns the WorldHotels, Best Western Hotels & Resorts, and | s1#20 s1#20 A: "three brands" became "the ... brands" in the body, the section and the record summary; the summary's narration of removed routes dropped. | |
| F14 ? | · | 2026-05-14/cve-2026-8043-ivanti-xtraction-external-file-control-cvss-9: Claim ledger 7069299494, 59dbcb21c1, b4ca050027. The cited Ivanti blog says only 'We have no evidence of these vulnerabilities b | s1#21 s1#21 A: "every flaw needs an authenticated account and none was exploited" replaced by Ivanti's own words (no evidence of exploitation in the wild) in the takeaway, sections, headline and summary. | |
| F6 strengthen-primary-source | · | 2026-05-13/certfr-2026-avi-0564-spip-4-4-14-multiple-rces-public-and-pr: (low confidence) The national-CERT restatement is listed as the primary and the vendor bulletin (https://blog.spip.net/Mise-a-j | s1#22 s1#22, s1#23 D: the first sources[] record stays the original primary under the entry lifecycle; both sourcing notes already name the vendor bulletin as the assessment. | |
| F6 strengthen-primary-source | · | 2026-05-13/certfr-2026-avi-0572-centreon-infra-monitoring-rce-sqli-xss: (low confidence) Same shape: the Centreon security bulletin on thewatch.centreon.com is the assessment per the sourcing_note but | s1#23 s1#22, s1#23 D: the first sources[] record stays the original primary under the entry lifecycle; both sourcing notes already name the vendor bulletin as the assessment. | |
| F7 drop | · | 2026-05-10/bauman-university-department-no-4-leaked-gru-cyber-operator: (low confidence) The entry's own takeaway states no defensive decision changes and the sources describe no attacker behavior (at | s1#24 s1#24 A: priority notable to routine (no defender decision, no attacker behavior). | |
| F7 drop | · | 2026-05-14/cve-2026-8043-ivanti-xtraction-external-file-control-cvss-9: (low confidence) Check 11: a vulnerability entry that demands no action beyond the regular patch cycle (authenticated flaws, ven | s1#25 s1#25 A: priority notable to routine. | |
| F8 needs-more-research | · | 2026-05-13/microsoft-mdash-multi-model-agentic-vulnerability-discovery: (low confidence) The cited MSRC CVE-2026-33824 advisory (MSRC API, mitigation article) says to block inbound traffic on UDP port | s1#26 s1#26 A: the takeaway adds MSRC's interim mitigation for CVE-2026-33824 (block or restrict inbound UDP 500 and 4500), cited to its advisory. | |
| F8 needs-more-research | · | 2026-05-13/mini-shai-hulud-s-github-actions-pwn-request-oidc-token-thef: (low confidence) The cited Wiz page's 2026-05-13 update says a payload bug renders the malware non-functional in the @uipath/* | s1#27 s1#27 A: Wiz's qualifier added: a payload bug leaves the malware non-functional in the @uipath and @mistralai packages. | |
| F9 surface-contradiction | · | 2026-05-11/bsi-flags-netgate-pfsense-community-edition-as-critical-unpa: (low confidence) The researcher's own CVSS vector is PR:L and the researcher's Full Disclosure post calls it 'expected behavior | s1#28 s1#28 A: the privilege disagreement is stated: Netgate disputes CVE-2025-69691 as admin-only, while the CVE record's vector scores low privileges (EUVD-2025-209739). | |
| F9 surface-contradiction | · | 2026-05-12/ico-fines-south-staffordshire-water-963-900-water-sector-oes: (low confidence) The MSRC API record gives baseScore 5.5 with vector AV:L/PR:L, which contradicts MSRC's own text (unauthentica | s1#29 s1#29 A: the CVE-2020-1472 cvss string now names MSRC's current 5.5 record and its disagreement with MSRC's own description. | |
| F10 missed-angle | · | run record: Coverage gap noticed: (low confidence; already tracked by the operator) The run record names the gap and sends it to the tracker. entries/2026-05-10/cpanel-whm-second-emergency-tsr... only | s1#30 s1#30 D: CVE-2026-41940 is already named as a coverage gap in the run record and goes to the tracker. | |
| F11 editorial-advisory | · | 2026-05-10/groupe-3r-r-seau-radiologique-romand-akira-ransomware-claims: Narrates removed content in the record summary (rendered on the site). | s1#31 s1#31 A: the record summary no longer narrates the removed playbook and actions. | |
| F11 editorial-advisory | · | 2026-05-10/bauman-university-department-no-4-leaked-gru-cyber-operator: Citation-housekeeping narration in reader-facing text. | s1#32 s1#32 A: the record summary states the VSquare fact instead of narrating citation housekeeping. | |
| F11 editorial-advisory | · | 2026-05-10/clickfix-campaign-expands-to-macos-macsync-shub-stealer-and: Both clauses narrate text the entry used to carry; state only what the sources say. | s1#33 s1#33 A: two "rather than" clauses narrating removed text dropped. | |
| F11 editorial-advisory | · | 2026-05-11/sms-blaster-smishing-establishing-itself-in-switzerland-port: Narrates removed text; no cited source carries the contrast. | s1#34 s1#34 A: "not an Android 12 network-type call" and "not through the TelephonyManager call" dropped. | |
| F11 editorial-advisory | · | 2026-05-12/bka-and-zit-dismantle-relaunched-crimenetwork-darknet-market: Narrates removed claims instead of stating a delta. | s1#35 s1#35 A: the clause narrating removed claims dropped. | |
| F11 editorial-advisory | · | 2026-05-12/ico-fines-south-staffordshire-water-963-900-water-sector-oes: Narrates removed text. | s1#36 s1#36 A: two "rather than" clauses narrating removed text dropped from the section and the record summary. | |
| F11 editorial-advisory | · | 2026-05-12/teampcp-unc6780-pcpjack-ecosystem-backdoors-the-checkmarx-je: Composition narration, uncited. | s1#37 s1#37 A: the composition sentence at the end of the section dropped. | |
| F11 editorial-advisory | · | 2026-05-13/microsoft-mdash-multi-model-agentic-vulnerability-discovery: Vendor product-efficacy figures (21/21, 96%/100% recall, CyberGym score) read as marketing; the entry labels them Microsoft's ow | s1#38 s1#38 D: the benchmark figures are attributed to Microsoft and describe the tool that is the entry's subject (advisory). | |
| F11 editorial-advisory | · | 2026-05-13/microsoft-mdash-multi-model-agentic-vulnerability-discovery: Composition narration. | s1#39 s1#39 A: "which is the exposure check now stated above" dropped. | |
| F11 editorial-advisory | · | 2026-05-13/trickmo-trickmo-c-android-banking-trojan-migrates-c2-to-the: Composition narration in reader-facing text (ledger 83e58a89c4 uncited). | s1#40 s1#40 A: composition narration removed from the section and the record summary. | |
| F11 editorial-advisory | · | 2026-05-14/dutch-igj-rules-clinical-diagnostics-nmdl-failed-nen-7510-in: Narration of edits in reader-facing text (ledger bb5b169871 uncited). | s1#41 s1#41 A: "that comparison is removed", "which is what the detection guidance above now rests on" and the priority narration dropped. | |
| F11 editorial-advisory | · | 2026-05-14/cve-2026-8043-ivanti-xtraction-external-file-control-cvss-9: Narrates removed text and what sources do not say; state only the corrected facts. | s1#42 s1#42 A: narration of removed claims dropped from the section. | |
| F3 claim-not-supported | · | Correction section 2026-10-21T20:28:35Z and record summary: RubyDoc.info "not a RubyGems service" (claims bd4bf64cde, 7964ec0658, de97ac9d83): (low confidence) The cited rubyhack.ai page says only tha | s2#1 s2#1 A: "not a RubyGems service" and "no source places RubyDoc.info inside RubyGems' trust boundary" removed from the section and the record summary (rubyhack.ai calls it RubyGems' automatic build system). | |
| F3 claim-not-supported | · | Exposure line (claim 5479e452f9): (low confidence) The RubyGems blog says "Gem installs and pushes for existing users remained unaffected" and found "no evidence that these attempts [key theft] succee | s2#2 s2#2 A: Exposure now says installs and pushes for existing users were unaffected (Ruby Central's words); "so no software consumed by others was compromised" dropped. | |
| F3 claim-not-supported | · | Triage line (claim db0c187d17): (low confidence) Socket only recommends blocking outbound gem push in CI that does not publish gems; it never says such a push has no benign explanation. | s2#3 s2#3 A: Triage clause replaced by Socket's own advice to block gem uploads from CI that does not publish. | |
| F3 claim-not-supported | · | 2026-09-19 section (claim 9c12f423ad): (low confidence) rubyhack.ai lists openaixyz65947@gmail.com under "Agents self-identified as being from OpenAI"; it does not say the address impersonates OpenAI. | s2#4 s2#4 A: "impersonating OpenAI" became "self-identifies with OpenAI" (rubyhack lists it under agents self-identified as OpenAI). | |
| F3 claim-not-supported | · | 2026-09-19 section (claim 64ebd61bd4): (low confidence) Euractiv says OpenAI "did notify a prior safety incident" involving Hugging Face and gives no month; the July date is on OpenAI's page, not the | s2#5 s2#5 A: "July's" dropped (Euractiv gives no month). | |
| F4 hallucinated-fact | · | 2026-09-19 section closing inference: (low confidence) Sources say the agents attempted the caching flaw ("we do not know if this attempt succeeded"; RubyGems "found no evidence that these attempts su | s2#6 s2#6 A: the "find and use unremediated flaws" inference dropped. | |
| F3 claim-not-supported | · | Body para 1 (claim d152caccf9): (low confidence) BankInfoSecurity says the group "regularly discussed ways to disable endpoint security tools ... modify Group Policy Objects and gain 'domain admin' pr | s2#7 s2#7 A: DynaRisk's statement is given as the group's discussions (BankInfoSecurity: "regularly discussed ways to disable endpoint security tools ... modify Group Policy Objects and gain domain admin"). | |
| F3 claim-not-supported | · | Defender takeaway (claim d149b34e3e): (low confidence) Bitdefender says generally that "actors will exploit and re-exploit the same access path until the original vulnerability is patched, compromised | s2#8 s2#8 A: the takeaway gives Bitdefender's general lesson (actors re-exploit the access path until it is patched and credentials rotated), not a claim about this case. | |
| F3 claim-not-supported | · | Body para 1 (claim 9c8973e09a): (low confidence) Datadog names the TanStack and UiPath compromises and lists artifacts router_init.js, setup.mjs and opensearch_init.js; it never calls out an OpenSearc | s2#9 s2#9 A: TanStack and UiPath, not OpenSearch. | |
| F14 ? | · | Body para 2 (claim e94901d5c6): (low confidence) Datadog: "OS-specific lists of over 100 sensitive file paths" and, in its ATT&CK table, "100+ file paths scanned across 3 OS families"; a per-OS count | s2#10 s2#10 A: OS-specific lists of more than 100 sensitive file paths. | |
| F13 ? | · | Correction section and record summary (claims 715ab07432, 98c1e90579): (low confidence) Datadog says a repository "noticed under several likely compromised users" appeared with what "appears to be the | s2#11 s2#11 A: "was not leaked" and "released deliberately" dropped from the section and the record summary; the README and commit facts stand. | |
| F3 claim-not-supported | · | Body para 1 (claim 09479b4d21): (low confidence) The BSI CSAF lists CVE-2026-45543, -45544, -45545 and -45691 only; it does not list CVE-2026-45690, -45156 or -45286 from the entry's batch. | s2#12 s2#12 A: BSI's WID-SEC-2026-1517 is said to cover the four CVEs its CSAF lists. | |
| F3 claim-not-supported | · | Defender takeaway (claim f7c1432725): (low confidence) AMD's table lists only OEM BIOS PI versions for Ryzen and is silent on OS updates; MSRC says "The mitigation for this vulnerability requires a Wi | s2#13 s2#13 A: the takeaway states what AMD lists (OEM BIOS PI versions for Ryzen-class parts) and what MSRC says (a Windows update for certain AMD models), not the inference that an OS update does not mitigate Ryzen… | |
| F4 hallucinated-fact | · | Headline (claim 8568f82b27): (low confidence) DEA: "Andresen allegedly accessed the Dream Wallets"; the headline states the reactivation as fact while the entry's own sourcing_note calls every wallet | s2#14 s2#14 A: headline attributes the wallet reactivation to US prosecutors. | |
| F4 hallucinated-fact | · | techniques[] T1597.002: (low confidence) T1597.002 (Search Closed Sources: Purchase Technical Data) names a reconnaissance behavior; DEA and The Record describe money laundering and an arrest, no ATT& | s2#15 s2#15 D: T1597.002 stays. The US Attorney states the marketplace earned commissions from selling stolen personally identifiable information, and T1597.002 is the purchase of victim data on cybercrime markets… | |
| F4 hallucinated-fact | · | techniques[] T1593.003: (low confidence) T1593.003 (Search Open Websites/Domains: Code Repositories) is an adversary behavior; Krebs reports no adversary searching repositories (GitGuardian is a defen | s2#16 s2#16 A: T1593.003 removed (the only repository scanning described is GitGuardian's, a defender). | |
| F3 claim-not-supported | · | cves[] cvss labels and correction section (claims e48483eb73, 6d0376822b, 90c0de34af, 7890d9f4e7, 4216d83ca8): (low confidence) Cyera gives 9.6/8.8/7.8/7.7 with no CVSS version, and the cited VulnChec | s2#17 s2#17 A: the "CVSS 3.1" label dropped from Cyera's scores in cves[] and the section. | |
| F3 claim-not-supported | · | GHSA-to-CVE binding (claims 84db452553, bcf05070c1): (low confidence) The GitHub page shows Severity Moderate but "CVE ID: No known CVE"; the pairing of each GHSA to a CVE id is by description only (E | s2#18 s2#18 A: the GHSA is described as the advisory for the heredoc allowlist bypass, not as carrying the CVE id. | |
| F13 ? | · | Body para 1 and cves[] (claims 5b8aeeab8b, e26091b786): (low confidence) SentinelOne only hyperlinks JetBrains' CVE-2023-42793 page on the word "TeamCity" and never names the CVE or the entry vector o | s2#19 s2#19 A: "JetBrains fixed that flaw" became "JetBrains fixed CVE-2023-42793 ..."; the body already says the post links JetBrains' advisory. | |
| F3 claim-not-supported | · | Body para 3 and summary (claim 2f7c5d26ff): (low confidence) The wFirma page says Java, the KIR plugin and Szafir Host are needed to sign with one's own qualified signature (Certum, KIR, Eurocert, Sim | s2#20 s2#20 A: "a Szafir qualified certificate" became "a qualified signature" in the body and summary. | |
| F4 hallucinated-fact | · | summary (claim 3ba49833fd): (low confidence) ESET: "The decision whether or not to deliver a payload is very likely manually performed by the operators"; the summary drops the hedge (the body and evid | s2#21 s2#21 A: ESET's hedge restored in the summary ("very likely decide manually"). | |
| F4 hallucinated-fact | · | summary (claim 0a651fc795): (low confidence) eSentire presents these as hunts and indicators and says its SOC isolated the customer accounts; it never says the campaign was first found through them. | s2#22 s2#22 A: summary says eSentire's hunts look for the signals, not that it found the campaign through them. | |
| F3 claim-not-supported | · | Correction section (claim d900e9737b): (low confidence) eSentire, cited throughout the entry, quotes the consent prompt "Are you trying to sign in to Microsoft Authentication Broker?" and describes th | s2#23 s2#23 A: the negative replaced by eSentire's quoted consent prompt. | |
| F3 claim-not-supported | · | Body para 1 (claim a76883459d): Security Affairs quotes the letter's April 8 sentence but never gives the letter date; the 2026-05-01 date is on the Cybersecurity Dive page ("letter sent to impacted i | s2#24 s2#24 A: the 2026-05-01 letter date is cited to Cybersecurity Dive. | |
| F3 claim-not-supported | · | Correction section (claim 3ee510cd1c): (low confidence) Sophos says "Attackers are increasingly using stolen credentials as their primary mechanism for delivery" and Help Net Security says the finding | s2#25 s2#25 A: the negative replaced by Sophos' own statement that attackers increasingly use stolen credentials as their primary delivery mechanism; the narration of removed cross-links dropped. | |
| F2 generic-url | · | KEV catalogue feed cited for per-CVE claims (also Nextcloud, Composer): (low confidence) The cited URL is the whole KEV catalogue feed (1,754 entries), not a per-CVE page; for the positive claim "list | s2#26 s2#26 D: KEV catalog feed (same reason as s1#1). | |
| F5 missing-citation | · | Correction section (claim 71259db24c): (low confidence) Uncited. The GHSA says "Classic ghp_ PATs are not affected by the regex bug per se, but the same leak primitive applies to any future credential | s2#27 s2#27 A: the uncited "Personal access tokens are not part of the issue" replaced by the advisory's own statement on classic ghp_ tokens, cited. | |
| F7 drop | · | Defender takeaway: "patch in the regular cycle" (also Nextcloud "upgrade in the regular cycle", F5 "in the regular cycle"): (low confidence) Check 11: a routine patch-cycle CVE with no exploitation an | s2#28 s2#28 A (this entry, AMD and F5): priority notable to routine for regular-cycle patches. | |
| F7 drop | · | No-nexus incident/news entries that state their own lack of defender impact (Interpol, BKA, Arwini, 7-Eleven, FunnelKit, SzafirHost): (low confidence) Check 5: breach/incident entries with no constitu | s2#29 s2#29 A (this entry, Interpol, Arwini, 7-Eleven): priority notable to routine. FunnelKit keeps notable (an exploited flaw with a patch). D: the bodies are not cut to two sentences, because entries are removed only… | |
| F12 single-source-flag-missing | · | verification: multi-source with sourcing_note "Composer's maintainers are the only assessor" (changed this run from single-source); also Sophos: The value contradicts the note on both entries (both fl | s2#30 s2#30 A (this entry and Sophos): verification multi-source to single-source, matching the sourcing notes. | |
| F16 ? | · | priority high on GemStuffer, CISA contractor leak and FamousSparrow: (low confidence) The `high` bar needs plausible exposure of this constituency and a decision within 7 days beyond the regular cycle | s2#31 s2#31 A (this entry): priority high to notable (no constituency exposure, no compromise of existing packages). | s2#31 A (this entry and the CISA contractor entry): priority high to notable. | |
| F17 ? | · | classification B/2: (low confidence) The entry's central claim (OpenAI agents behind GemStuffer, RubyDoc.info RCE) rests on one researcher group that RubyGems says it cannot confirm and OpenAI says it | s2#32 s2#32 D: credibility 2 stays; the entry's core finding (the registry used as an exfiltration channel) rests on Socket's analysis and RubyGems' own account, two assessors; the attribution is presented as the… | |
| F8 needs-more-research | · | Detection/body omit a documented behavior: (low confidence) The source documents a workflow-injection secret-dump path and its cleanup that a hunter could look for in Actions audit logs; the entry's D | s2#33 s2#33 A: Detection adds the disguised secrets-dumping workflow and the deleted workflow runs and branches (Datadog's table). | |
| F6 strengthen-primary-source | · | sources[] order and F5 primary: (low confidence) Put eSentire first on the Tycoon entry. On the F5 entry the vendor primary could not be read by any rung; the facts are carried by SecurityWeek and EUV | s2#34 s2#34 D: the first sources[] record stays the original primary (lifecycle); for F5, SecurityWeek and EUVD carry the cited facts and the sourcing note says so. | |
| F9 surface-contradiction | · | Charter statement vs later leak on the cited page: (low confidence) The entry (summary, 2026-05-25 section, correction) cites CyberInsider for Charter's denial that "sensitive personal information (PI | s2#35 s2#35 A: this run's section adds CyberInsider's later update (Charter dataset leaked, 4.9 million email addresses per Have I Been Pwned). | |
| F9 surface-contradiction | · | Advisory prose vs rating: (low confidence) The entry reports only "High 7.5"; the advisory's own description says medium. A one-line Contradiction note would surface it. | s2#36 s2#36 A: the advisory's High 7.5 header and its "Severity: medium" text are both stated. | |
| F11 editorial-advisory | · | Record-keeping narration in reader-facing sections and record summaries: These narrate what the entry used to say or how it was edited; per check 12 they are workflow-internal composition rationale in | s2#37 s2#37 A (this entry): "replace detection advice Socket did not give" removed from section and summary. | s2#37 A (this entry): "The detection guidance now follows ..." dropped. | s2#37 A: narration of "the earlier… | |
| F11 editorial-advisory | · | (pre-existing, append-only records) stale and non-compliant record summaries: The records cannot be edited, but they still assert what the corrected sections now disclaim ("never filed" vs "as of 2026 | s2#38 s2#38 D: earlier records' summaries are append-only (2026-09-19 GemStuffer, 2026-05-25 7-Eleven). | |
| F11 editorial-advisory | · | tags contradict the correction: The correction says the packages were not designed for developer compromise and existing users were unaffected, and the scraped council data was publicly accessible; th | s2#39 s2#39 A: tags supply-chain and data-breach removed. | |
| F11 editorial-advisory | · | Fetch-failure narration in sourcing_note: Provenance in two sentences is fine, but how a page failed to render is operator narration; state which source carries the fact. | s2#40 s2#40 A: F5 and 7-Eleven sourcing notes state which source carries the facts, without narrating fetch failures. | |
| F11 editorial-advisory | · | Registry summary for tool:gemstuffer-rubygems-2026 is stale: The sibling incident record was rewritten but this summary still carries the pre-correction framing ("weaponised", "inbound-monitoring blin | s2#41 s2#41 A: registry summary of tool:gemstuffer-rubygems-2026 rewritten. | |
| F11 editorial-advisory | · | Attacker-chosen artifact names in Detection (file-name indicators): Check 12 lists file-name indicators among IOCs. These are tool-specific names rather than hashes or domains and the guidance is beha | s2#42 s2#42 D: tool and loader file names (LMIGuardianSvc.exe, winmm.dll, gh-token-monitor, setup.mjs) are behavioral triage artifacts, not hashes, IPs or domains (advisory). | |
| F11 editorial-advisory | · | Awkward version strings: Repeated "version" prefix reads as a transcription artifact; "31.0.14.5, 30.0.17.9 and 29.0.16.16" is clearer. Also evidence quote typography: gemstuffer quote begins "We temp | s2#43 s2#43 A (quote case): evidence quote starts with lowercase "we" as on the page. | s2#43 D: "version" stays before four-part version numbers so the IOC scan does not read them as IPv4 addresses. | |
| F3 claim-not-supported | · | Correction 2026-10-21T20:43:20Z (claim cfe8e26f82): PwC says JFMBackdoor supports "remote shell access, file system operations, network proxying, screenshot capture, and self-removal capabilities" and | s3#1 s3#1 A: Calypso, JFMBackdoor's capabilities follow PwC, self-removal included. | |
| F3 claim-not-supported | · | Body, Azure control plane paragraph (claim 3695d229cb): Microsoft hedges: "Among these secrets, we believe the threat actor found credentials that enabled them to access the application they coveted t | s3#2 s3#2 A: Storm-2949, the takeover of the web app is Microsoft's belief. | |
| F3 claim-not-supported | · | Body, Virtual machines paragraph (3350cc7f83) and Correction (1a10bf5851): (low confidence) Microsoft writes "The script attempted to weaken Microsoft Defender Antivirus by disabling several protectio | s3#3 s3#3 A: Storm-2949, the Defender step is an attempt ("tried to weaken"); the log clearing stays (Microsoft's detections table lists it). | |
| F3 claim-not-supported | · | Body, first sentence (claim f9d1fe1220): BleepingComputer says only "It received the maximum severity score from HiddenLayer, the company that discovered it"; neither "ChromaToast" nor "CVSS 4.0" appe | s3#4 s3#4 A: ChromaDB, "ChromaToast" and the CVSS 4.0 label are cited to Hadrian, BleepingComputer keeps the maximum-severity clause. | |
| F3 claim-not-supported | · | Correction 2026-10-21T20:41:34Z (claim 025eb49c6a): (low confidence) The ESET page links Webworm to SixLittleMonkeys and FishMonger but never says "Aquatic Panda"; that alias is on the Hacker News pag | s3#5 s3#5 A: Webworm, the Aquatic Panda alias is cited to The Hacker News. | |
| F4 hallucinated-fact | · | summary (claim e34b516f34): (low confidence) "them" = CVE-2026-7507 and CVE-2026-37982. RHSA-2026:19594 (26.2.16) lists CVE-2026-7507, -7307 and -7504 only; CVE-2026-37982 is only in RHSA-2026:19596 ( | s3#6 s3#6 A: Keycloak, the summary splits the fixes (CVE-2026-7507 in 26.2.16 and 26.4.12, CVE-2026-37982 in 26.4.12). | |
| F3 claim-not-supported | · | Body, first paragraph (claim 6ef078714c): (low confidence) THN says the exploits were "disclosed by Chaotic Eclipse ... last month"; 2026-04-16 is in Help Net Security (cited elsewhere). THN's confirm | s3#7 s3#7 A: CVE-2026-41091, the 2026-04-16 publication is cited to Help Net Security. | |
| F3 claim-not-supported | · | Body, Detection (claim b048a88e92): (low confidence) Help Net Security: "The attacker dropped the exploit files into the user's Pictures and Downloads folders and renamed them ... Then, before launchi | s3#8 s3#8 A: CVE-2026-41091, the Detection order follows Help Net Security (files dropped and renamed, then discovery, then the exploits). | |
| F3 claim-not-supported | · | Body (bd8345aaf8) and Correction (0cb1791527): (low confidence) The CCB advisory lists the CVE/CVSS values and vectors but never calls them the CNA's scores; the ENISA EUVD records (assigner CERT-PL, | s3#9 s3#9, s3#16 A: Sparx, the scores are attributed to CCB and the disagreement with the vendor is stated. | |
| F3 claim-not-supported | · | Body, Exposure (claim cae6cfdffb): (low confidence) SA-CORE-2026-004: "The Drupal releases for supported branches (11.3, 11.2, 10.6, and 10.5) in this advisory also include security updates for Symfon | s3#10 s3#10, s3#11 A: Drupal, Exposure and Detection follow the advisories (Symfony/Twig wording, the scanning markers). | |
| F14 ? | · | Body, Detection (claim 3d399370e4): (low confidence) Imperva: "Several payloads include Nuclei-style markers". | s3#11 s3#10, s3#11 A: Drupal, Exposure and Detection follow the advisories (Symfony/Twig wording, the scanning markers). | |
| F4 hallucinated-fact | · | cves[CVE-2026-26956] (claim 033a18ab7d): (low confidence) GHSA-ffh4-j6h5-pg66: "Affected versions <= 3.10.4, Patched versions 3.10.5"; the record narrows the range to a single version (THN wording) an | s3#12 s3#12 A: vm2, CVE-2026-26956 affected is "3.10.4 and earlier" per the GHSA. | |
| F4 hallucinated-fact | · | cves[CVE-2026-44007] (claim e10aaaea45): (low confidence) GHSA-8hg8-63c5-gwmx vector CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:H (privileges required High, score 9.1); the same entry labels PR:L CVE- | s3#13 s3#13 A: vm2, CVE-2026-44007 auth is admin-required (PR:H). | |
| F4 hallucinated-fact | · | techniques[] T1014: (low confidence) T1014 is Rootkit (hiding programs, files, services, drivers); no cited source describes hiding behavior. Broadcom describes a filesystem filter driver that patches | s3#14 s3#14 A: fast16, T1014 removed. | |
| F14 ? | · | title: (low confidence) V12/oss-security: "only default on Arch Linux among the common distributions we tested"; the entry's own Improvement says V12 "did not establish that the module is off on every | s3#15 s3#15 A: PinTheft, the title says "among the distributions it tested". | |
| F9 surface-contradiction | · | source CCB advisory: (low confidence) The cited advisory's title says actively exploited, its body says no proof of exploitation; the entry silently follows the body (correctly) with no Contradiction | s3#16 s3#9, s3#16 A: Sparx, the scores are attributed to CCB and the disagreement with the vendor is stated. | |
| F9 surface-contradiction | · | StepSecurity vs Nx timeline: (low confidence) Two cited sources disagree on the publish time; the entry takes Nx's times without a Contradiction note. | s3#17 s3#17 A: Nx, StepSecurity's 12:36 UTC publication time is stated beside Nx's times. | |
| F9 surface-contradiction | · | HiddenLayer contact attempts: (low confidence) Hadrian (also cited): HiddenLayer "got an initial acknowledgement on 2025-12-16 followed by silence across 10+ follow-ups". The entry notes the two repor | s3#18 s3#18 A: ChromaDB, Hadrian's acknowledgement on 2025-12-16 and the silence after it are stated. | |
| F8 needs-more-research | · | Detection: (low confidence) The same Learn page puts combined MFA registration events under Security > Authentication Methods; the attacker's new Authenticator registration is found there, which Detec | s3#19 s3#19 A: Storm-2949, Detection adds the Authentication Methods registration events. | |
| F8 needs-more-research | · | affected_products / cves[].affected / fixed: (low confidence) The cited advisory also lists SMA100 10.2.1.13-72sv and older (fixed 10.2.1.14-75sv), Gen7 7.0.1-5161 / 7.1.1-7058 lines (fixed 7.0.1-5169 | s3#20 s3#20 A: SonicWall, cves[] affected and fixed carry the full advisory table (Gen7 7.0.1 and 7.1.1 lines, Gen8 TZ80, Gen6 NSv fix, SMA100), Exposure adds one sentence, and SonicWall SMA 100 is in affected_products. | |
| F8 needs-more-research | · | cves[CVE-2026-6856] cvss: (low confidence) Red Hat's CVE-2026-6856 page carries a preliminary Red Hat base score 3.1 (CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:N/I:L/A:N); the entry cites only the GitHub iss | s3#21 s3#21 D: Keycloak CVE-2026-6856 keeps no score: Red Hat's figure is a preliminary score on a page the entry does not cite, and the CVE's facts are carried by the GitHub issue. | |
| F7 drop | · | whole entry (notable): (low confidence) Unexploited, not publicly disclosed at release, auto-updating engine flaw needing multiple user actions; a routine patch-cycle CVE with no exposure-driven urgen | s3#22 s3#22, s3#23, s3#24 A: Defender CVE-2026-45584, Azure Local CVE-2026-42822 and B1ack's Stash moved from notable to routine. | |
| F7 drop | · | whole entry (notable): (low confidence) Unexploited CVE in a restricted allow-listed offering with an internal-network/insider precondition; the entry says itself there is no decision beyond the plann | s3#23 s3#22, s3#23, s3#24 A: Defender CVE-2026-45584, Azure Local CVE-2026-42822 and B1ack's Stash moved from notable to routine. | |
| F7 drop | · | whole entry (notable): (low confidence) Incident entry with no Swiss/EU-public-sector nexus (about 70% US cardholders, Europe only through UK and French cards) and no TTP beyond a marketplace announce | s3#24 s3#22, s3#23, s3#24 A: Defender CVE-2026-45584, Azure Local CVE-2026-42822 and B1ack's Stash moved from notable to routine. | |
| F16 ? | · | priority: high: (low confidence) Asia/Middle East telecom espionage (Afghanistan ISP, Azerbaijan, Kazakhstan/India targeting); no plausible exposure of the Swiss public-sector constituency is shown, s | s3#25 s3#25 A: Calypso, high to notable. | |
| F11 editorial-advisory | · | Correction/Improvement sections retell the retracted text: The bodies were rewritten wholesale, so the sections mostly narrate what the entry used to say, which a reader never saw (check 4c(i)). The g | s3#26 s3#26 D: the correction sections state what the published entries got wrong; those entries were live since May, so readers saw the statements now corrected (docs/pipeline.md, record types: a correction section… | |
| F11 editorial-advisory | · | updates[].summary: Record summaries render on the site; these clauses narrate the rewrite process and field changes rather than what the section tells a reader (pipeline self-reference). | s3#27 s3#27 A: 17 record summaries rewritten to state the finding rather than the rewrite (main/narrating-summaries.json). | |
| F11 editorial-advisory | · | summary, body and 2026-05-22 Update: MSRC gives a CVSS 3.1 base score of 4.0 (vector CVSS:3.1/AV:L/AC:L/PR:N/...); in prose "CVSS 4.0" reads as CVSS version 4.0 (the Sparx entry uses that form for v4 | s3#28 s3#28 A: CVE-2026-41091, "CVSS 3.1 base score 4.0". | |
| F11 editorial-advisory | · | body: (low confidence) Body is a run of year-over-year percentages with a one-sentence takeaway (check 7); the DBIR is primary research and priority is notable, so advisory only. | s3#29 s3#29 D: the DBIR entry is an annual report whose content is the year-over-year figures, kept at notable as primary research. | |
| F11 editorial-advisory | · | sources[] bk.admin.ch: (low confidence) The Federal Chancellery page says the federal standard web platform is based on Livingdocs; it supports only a negative nexus remark in the Correction, not the | s3#30 s3#30 D: sources[] roles are only primary and corroborating; the Chancellery page backs the entry's Swiss exposure remark and stays. | |
| F11 editorial-advisory | · | "Together they map 400 ATT&CK ids, all active" / "400 active ATT&CK ids are mapped": Counting techniques[] across the 74 updated entries gives 400 mappings but 173 distinct ids; all are active in the | s3#31 s3#31 A: report and run record give 398 mappings over 171 distinct ids (recounted after the slice fixes) and 35 priority moves. | |
| F11 editorial-advisory | · | campaign:calypso-red-lamassu-showboat-jfmbackdoor-linux-windows-telco (pre-existing): (low confidence, pre-existing) Neither Lumen nor PwC, the entry's sources, uses the name Bronze Medley; the record | s3#32 s3#32 D (pre-existing): the Bronze Medley alias on the Calypso campaign record goes to tracker row W4-07 with the Red Lamassu actor proposal. |
Iteration #? NEEDS_FIXES cap-breach · 16 findings (truth=9, editorial=3, advisory=4) · Claude Sonnet 5.5 · 28m 52s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F4 hallucinated-fact | · | 2026-05-14/cve-2026-8043-ivanti-xtraction-external-file-control-cvss-9 (claim 70fbc86db6): The Ivanti entry's record summary is a verbatim copy of the TrickMo entry's record summary (same text in entr | i2#1 i2#1 A: Ivanti record summary replaced with the Ivanti correction text (the TrickMo text had been pasted in by the s1#42 fix). | |
| F3 claim-not-supported | · | 2026-05-22/calypso-red-lamassu-bronze-medley-deploys-showboat-linux-and (claim 5104ea6f4e): The cited PwC page carries 'since 2019' and the JFMBackdoor capabilities but no 2022 date (grep of the page: | i2#2 i2#2 A: Calypso, "at least mid-2022" cited to Lumen. | |
| F3 claim-not-supported | · | 2026-05-13/bwh-hotels-best-western-worldhotels-sure-hotels-181-day-unau (claims ed80615e6a, d0aa77e5de): (low confidence) The Register (read with WebFetch this pass) says 'No payment or bank details w | i2#3 i2#3 A: BWH, "passport" dropped from the body and the record summary (The Register says no payment or bank details). | |
| F3 claim-not-supported | · | 2026-05-16/amd-sb-7052-cve-2025-54518-amd-zen-2-op-cache-corruption-soc (claim c0cb72e3a0; frontmatter summary): (low confidence) AMD's table says 'EPYC 7002 ... OS Update | Contact your OS Vendor'; M | i2#4 i2#4 A: AMD, the frontmatter summary and record summary no longer bind the Windows mitigation to EPYC 7002; Microsoft says the latest Windows builds enable a mitigation for the affected AMD models without naming… | |
| F3 claim-not-supported | · | 2026-05-21/cve-2026-45829-chromadb-python-fastapi-server-pre-auth-rce-v (claim 14627a4b5b): (low confidence) Hadrian writes 'The bug was reported on 2025-11-28 by HiddenLayer ("ChromaToast")' and link | i2#5 i2#5 A: ChromaDB, Hadrian repeats HiddenLayer's name "ChromaToast". | |
| F3 claim-not-supported | · | 2026-05-20/drupal-core-highly-critical-pre-patch-warning-unauthenticate (claim 80f1e9f3db): (low confidence) SA-CORE-2026-004: 'The Drupal releases for supported branches (11.3, 11.2, 10.6, and 10.5) | i2#6 i2#6 A: Drupal, the Symfony and Twig updates follow SA-CORE-2026-004 (highly recommended whether or not the SQL injection applies, exposure depends on configuration and contributed modules). | |
| F4 hallucinated-fact | · | run record 'Scope and outcome' and wave report § Registry and state changes / summary: On disk the tallies are 399 mappings (171 distinct; the main agent's own work/.../main/stats.py prints 399 171), | i2#7 i2#7 A: run record and report recounted: 398 mappings over 170 distinct ids (after i2#9), 23 CVEs added, 40 product entities, 40 summaries, 36 priority moves (after i2#12), three primary URLs. | |
| F4 hallucinated-fact | · | docs/audits/2026-10-21-cleanup-wave-04.md per-entry table rows for cPanel, BWH, Datadog, Sparx: Six table rows still carry the pre-remediation record text (cPanel, BWH, Datadog, Sparx, AMD, Ivanti). F | i2#8 i2#8 A: the report's per-entry table regenerated from the final record summaries. | |
| F4 hallucinated-fact | · | 2026-05-14/gemstuffer-rubygems-weaponised-as-a-one-way-exfiltration-cha techniques[] T1567.004: (low confidence) T1567.004 in the pinned dataset is 'Exfiltration Over Webhook'. No cited source (Socket | i2#9 i2#7 A: run record and report recounted: 398 mappings over 170 distinct ids (after i2#9), 23 CVEs added, 40 product entities, 40 summaries, 36 priority moves (after i2#12), three primary URLs. | i2#9 A: GemStuffer… | |
| F5 missing-citation | · | 2026-05-12/ico-fines-south-staffordshire-water-963-900-water-sector-oes cves[CVE-2020-1472].status: (low confidence) 'cisa-kev' is set with no KEV source in sources[] or the body (the cached KEV feed | i2#10 i2#10 A: ICO, cisa-kev dropped from CVE-2020-1472's status (no KEV source cited in the entry). | |
| F12 single-source-flag-missing | · | 2026-05-17/cve-2026-41225-f5-big-ip-big-iq-icontrol-rest-manager-role-a: (low confidence) The note states F5 is the only assessor and the other sources restate it; s2#30 flipped Composer and Sophos to | i2#11 i2#11 A: F5 CVE-2026-41225, verification single-source (F5 is the only assessor). | |
| F16 ? | · | 2026-05-10/groupe-3r-r-seau-radiologique-romand-akira-ransomware-claims priority: high: (low confidence) A Swiss in-region incident clears the exposure half of the high bar, but the entry itself says | i2#12 i2#7 A: run record and report recounted: 398 mappings over 170 distinct ids (after i2#9), 23 CVEs added, 40 product entities, 40 summaries, 36 priority moves (after i2#12), three primary URLs. | i2#12 A: Groupe… | |
| F11 editorial-advisory | · | 2026-05-21/cve-2026-45829-chromadb-python-fastapi-server-pre-auth-rce-v record summary: Record-keeping narration of the edit in a site-rendered record summary. s3#27 lists cve-2026-45829 among the 17 | i2#13 i2#13 A: ChromaDB record summary ends on the sourced roles, no narration. | Sweep after i2#13/14: three more record summaries narrated pipeline internals and were tightened (West Pharmaceutical "only encryption is… | |
| F11 editorial-advisory | · | 2026-05-21/verizon-2026-dbir-vulnerability-exploitation-overtakes-crede record summary: Pipeline self-reference (source swap, priority and deep-dive change) in a reader-facing record summary; s3#27 li | i2#14 i2#14 A: Verizon record summary ends on the methodology change, no narration of the source swap or priority. | |
| F11 editorial-advisory | · | 2026-05-14/cve-2026-8043-ivanti-xtraction-external-file-control-cvss-9 sourcing_note: Fetch-failure narration in sourcing_note, the pattern s2#40 removed from the F5 and 7-Eleven notes. State which so | i2#15 i2#15 A: Ivanti sourcing_note, the fetch-failure sentence dropped. | |
| F11 editorial-advisory | · | (pre-existing) 2026-05-12/gtig-ai-threat-tracker-may-2026-first-confirmed-ai-generated Defender takeaway: (low confidence, pre-existing) GTIG says SANDCLOCK extracts 'cloud secrets, such as AWS keys a | i2#16 i2#16 A (pre-existing): GTIG takeaway, AWS keys and GitHub tokens are treated as stolen and the AI API secrets are those GTIG says the LiteLLM compromise could expose. |
Verification & coverage notes
The run record's narrative body, verbatim. This is where the run accounts for its own judgement calls: every borderline drop and judged-not-relevant item with its reason, dedup decisions, single-source items and their carve-outs, contradictions, and per-source coverage gaps, so nothing the run considered disappears silently.
Verification & coverage notesrun record body
2026-10-21T1148Z-audit · audit · Opus 5.5 · 0 entries published
Verification & coverage notes
Cleanup wave 04 (lane B, row W2-01) of the one-time store cleanup program (prompts/cleanup-coordinator.md), an operator-directed partial audit with the claim partial 2026-10-21T11:48:06Z wave=04; it ran no coverage re-sweep, so it does not move the weekly audit's window anchor. Report: docs/audits/2026-10-21-cleanup-wave-04.md. Tracker: docs/cleanup/STATUS.md.
Takeover. A first session claimed the wave (bca18bd) in an environment without the right network access and stopped after the claim. This session adopted that claim and its run id, re-synced main (base 7e74158) and pushed the planning state on its own (5a2e81b on main).
Scope and outcome. W2-01 listed 76 legacy entries filed 2026-05-10 to 2026-05-22; one had been folded by the 2026-10-11 audit, so 75 were truth-checked whole and none was clean (69 factual-error, 6 imprecision; 1,076 ledger claims, 405 held). 74 entries carry one record from this fire (68 corrections, 6 improvements); the 75th, the Drupal "emergency patch window" entry, was folded into the Drupal PSA-2026-05-18 entry, which closes tracker row W4-09. Every entry now has an Admiralty rating and a Defender takeaway, 398 technique mappings over 170 distinct active ATT&CK ids are in place, actions[] is empty on all 74, and 36 priorities moved down. The legacy queue went from 406 to 330 pending.
State. Registry: 7 new entities, 40 summaries rewritten, 12 relations, aliases moved where an actor name sat on a campaign record, and 40 product entities from tools/sync_products.py. cves_seen: 23 CVEs added, last_seen moved on the wave's records, titles and three primaries corrected. Cross-wave registry items (Webworm aliases, actor:kimsuky, actor:akira, campaign:clickfix-macos-2026, the Nightmare Eclipse campaign, a Red Lamassu actor) went to the tracker rather than being edited in passing.
Warnings left in run scope, each explained. aggregator-only on the Braintrust entry: its incident notice sits on a trust-center page that renders only in a browser (the bridge returns an empty shell, WebFetch only the title), so the entry rests on TechCrunch and SecurityWeek, which quote the email and the notice, with reduced confidence stated. attack-mapping on two research entries (NCSC UK AI guidance, Bauman University leak) whose sources describe no attacker behavior; the Sophos survey maps Valid Accounts from Sophos' own statement on stolen credentials. check_run.py --all shows no warning from this wave.
Budget and sub-agents. The session started before the repository was attached, so .claude/agents/ never registered: on the operator's interactive decision the four truth passes ran as claude agents pinned to Sonnet at xhigh effort that loaded .claude/agents/cti-verification.md as their definition (a deviation from the named-verifier rule, disclosed here and in the report). Two account-wide session-limit stops (HTTP 429, request ids in work/2026-10-21T1148Z-audit/main/spawn-notes.md) halted all running agents; after each reset the operator said to continue and each agent was resumed once, which the budget counts as a spawn. The resumed truth passes ran on Opus 5.5. 16 of 20 spawns were used before the verifier loop (4 truth passes, 4 repair workers and 8 continuations) and 4 in it (three first-pass slices and one post-fix pass), 20 of 20 in all. The first legs reported no usage, so the token sum covers the continuations and the verifier passes: about 7.66 M (truth 2.77 M, repairs 2.65 M, verifier 2.24 M).
Duration. duration_seconds is 89,772 (24.9 h), over the 24 h stall threshold, and the run did not hang. The claim was written at 11:48Z on 2026-10-21. Work stopped twice on the account session limit, at about 12:2xZ and 16:3xZ, and resumed after each reset. The session then sat idle from 21:02Z until the operator resumed it at about 11:00Z on 2026-10-22. The verifier loop ran that morning. The run-record WARN on the duration is this run's own telemetry and stays for the next audit to acknowledge as settled history.
Coverage gap noticed. CVE-2026-41940 (cPanel & WHM authentication bypass, KEV 2026-04-30, known ransomware use) has no entry of its own; tracker row W6-08.
Verification. Two iterations of the pinned cti-verification definition (Sonnet 5.5). Iteration 1, the first pass in three slices, answered all 1,535 claims this fire changed and returned NEEDS_FIXES (62 truth, 29 editorial, 26 advisory). The main agent applied 101 findings, two of them in part, and declined 16 with a rebuttal each in work/2026-10-21T1148Z-audit/main/iter1-remediation.md. The correction sections state what the published May entries got wrong, which readers saw (s3#26). Earlier records' summaries are append-only (s2#38). The KEV catalog feed stays the KEV citation, since CISA publishes no per-CVE page (s1#1, s2#26). The first sources[] record stays the original primary (s1#22, s1#23, s2#34). T1597.002 stays on Dream Market as the only behavior its sources state, and the kind question went to the operator as W7-18 (s2#15). The rest are kept on their source reading (s1#38, s2#32, s2#42, s2#43, s3#21, s3#29, s3#30) or sent to the tracker (s1#30 as W6-08, s3#32 to W4-07).
Iteration 2, the post-fix pass, answered the 135 claims changed since iteration 1 and re-checked the 101 applied fixes (96 held, 5 failed), returning NEEDS_FIXES (9 truth, 3 editorial, 4 advisory). The firmest was an Ivanti record summary that the s1#42 fix had overwritten with the TrickMo text. The others were citation bindings (Calypso, BWH, ChromaDB, Drupal, AMD), stale counts and table rows in this record and the report, an unsupported T1567.004 on GemStuffer, a cisa-kev status without a KEV source, a single-assessor entry marked multi-source and the Groupe 3R priority. All 16 were applied (work/2026-10-21T1148Z-audit/main/iter2-remediation.md), and a sweep tightened three more record summaries.
That pass was the wave's 20th spawn, so it covered the changed claims rather than the full post-fix scope of about 1,260 claims, and no confirmation pass could follow. The loop ended on the fail-open path with verification.budget_cut. The residual of 12 is iteration 2's truth plus editorial count, all applied and none re-verified.
← Operations dashboard · day page 2026-10-21 · run-record contract: docs/pipeline.md