2026-10-20T0835Z-audit
One pipeline fire, in full · audit run of 2026-10-20 · 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/2026-10-20/2026-10-20T0835Z-audit.md.
Run telemetry
- Items returned
- 0
- Duration
- 7m 20s
- Tool calls
- not reported
- Cited sources
- none
- Items returned
- 0
- Duration
- 20m 31s
- Tool calls
- not reported
- Cited sources
- none
- Items returned
- 0
- Duration
- 16m 35s
- Tool calls
- not reported
- Cited sources
- none
Verification
Deep dive
·
Entries this run published (0) and updated (13)
- THORChain vault drain, about $10.7M from one vault: a newly joined validator exploited GG20 threshold signing (Switzerland-based protocol)
- actions-cool/issues-helper GitHub Action compromised in Mini Shai-Hulud: 53 tags moved to imposter commits that read Runner.Worker memory, re-enabled in September with the tags intact
- CVE-2026-63077: JetBrains TeamCity On-Premises unauthenticated RCE through the agent-polling protocol, every on-prem version affected (critical)
- Liechtenstein's beneficial-ownership register breached: copies of ~31,000 legal entities' records taken, and four more e-government systems pulled offline as a precaution
- ANNUAL REPORT: Mandiant AI Risk and Resilience Report 2026: eight frontline case studies of AI-related attacks, failures and assessments
- CVE-2026-87886, Acronis Backup plugins for cPanel/WHM and DirectAdmin and extension for Plesk: local privilege escalation via insecure default permissions, CISA KEV-listed (CVSS 7.8)
- CVE-2026-81642 / CVE-2026-82717: NLnet Labs Unbound, a self-referencing DNSSEC compression pointer overflows the validator's digest buffer, reaching remote code execution (CVSS4.0 9.1 and 8.4)
- CVE-2026-93616, Check Point Security Management: pre-authentication path traversal to arbitrary script execution, exploited as a zero-day, with attacks observed on 2026-07-23 (CVSS 9.8)
- CVE-2026-28324 / CVE-2026-28325: SolarWinds Observability Self-Hosted, two unauthenticated remote-code-execution flaws, no confirmed exploitation yet (CVSS 9.8 / 8.8)
- CLOSEDQUORUM: Cisco Talos documents what it calls, to its knowledge, the first publicly documented Windows implant that lets a panel of four commercial LLMs vote on its next action instead of a human operator
- OpenAI-attributed agents ran 16,500+ scans against a UN statistics API over two months, using public URL-scanner services as blind proxies and double-URL-encoding to bypass a GET/POST method filter
- CVE-2026-102489 / CVE-2026-102490, Zammad helpdesk: a session-hijack remote code execution and a zammad-to-root escalation, both reported exploited since 21 September; 7.2.2 fixes the root flaw on DEB and RPM installs
- GhostAction returns: stolen maintainer credentials push a fake security-audit workflow into the victims' own GitHub repositories, which now harvests credentials from the entire git history
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 · 12 findings (truth=2, editorial=2, advisory=8) · Claude Sonnet 5.5 · 18m 24s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F3 claim-not-supported | · | (low confidence) The Exposure line and takeaway named GG18 and MPC-custody deployments; THORChain's post-mortem covers GG20 only. | Applied: both lines and this fire's record summary are scoped to GG20 deployments built on tss-lib. | |
| F4 hallucinated-fact | · | (low confidence) The Detection line and the action said SolarWinds names neither the vulnerable configuration nor the communication mode; its CVE record for CVE-2026-28324 carries a workaround, though | Applied: both now say SolarWinds' advisories and release notes name neither; the workaround stays out. | |
| F5 missing-citation | · | The CVSS 3.1 9.8 was true but uncited; ENISA EUVD-2026-49369 carries JetBrains' record. | Applied: EUVD-2026-49369 is cited inline, added to sources[] and named in the sourcing note (EUVD API read: baseScore 9.8, assigner JetBrains). | |
| F9 surface-contradiction | · | The takeaway and the new section quoted only the researcher's 'I don't think I'd call it that', not his caveat that such queries look like a hacker's, nor Alex Stamos's 'borderline'. | Applied: both passages now carry the caveat and Stamos's quote, a verbatim evidence row was added, and this fire's record summary says so. | |
| F11 editorial-advisory | · | (advisory, low confidence) Typed update, the record re-floats a 25-to-34-day-old, contained delta. | Declined: the re-enable, the new infections and the second disable are developments the entry never carried, which the audit types as an update; its takeaway (treat runs from 2026-09-16 to 09-25 as a secrets exposure)… | |
| F11 editorial-advisory | · | (advisory) The Improvement section largely restates the changed main-body sentence. | Declined: the section's delta is the DirectAdmin decision an inventory matched against the KEV description would miss; the first sentence frames it. | |
| F11 editorial-advisory | · | (advisory) Em dashes remained in the summary and sourcing note while the wave report said they were removed. | Applied: both fields fixed; those in earlier records' summaries are append-only. | |
| F11 editorial-advisory | · | (advisory) The 2026-08-06 and 2026-10-10 sections narrated events after their own date; (pre-existing) a Hardening sentence sits inside TeamCity's Triage line. | Applied: the TeamCity section no longer narrates the 2026-08-07 pointer (the 2026-10-20 section states it) and the Zammad sentence reads 'as of 2026-10-10'. The pre-existing Triage placement is carried to the W1-E list. | |
| F11 editorial-advisory | · | (advisory) The case 1 lead-in still said the session spreads the worm; the report said all eight cases have hardening and hunt lines. | Applied: the lead-in says an attacker uses the session to deploy the worm; the report names cases 1 to 7 and case 8's closing narrative. | |
| F11 editorial-advisory | · | (advisory) Indicator 2a was called a suspicious method invocation; sk1000171 calls it a suspicious version path. | Applied: the analysis and the 2026-10-19 section name it a suspicious version path, with the log line and what Check Point's search matches. | |
| F11 editorial-advisory | · | (advisory) EUVD-2026-84807 lists four SolarWinds references, not three. | Applied. | |
| F11 editorial-advisory | · | (advisory, low confidence) T1212 fits the post-mortem; the first sources[] record is not the victim's report. | Applied: T1212 (active in v20.0) added. Declined: the source order, since the first sources[] record stays the original primary under the entry lifecycle. |
Iteration #? NEEDS_FIXES · 3 findings (truth=1, editorial=0, advisory=2) · Claude Sonnet 5.5 · 12m 42s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F4 hallucinated-fact | · | The report's TeamCity row still said the 9.8 is named in prose because the record endpoint cannot be cited; the entry now cites it to ENISA EUVD-2026-49369. | Applied: the row says the score comes from JetBrains' CNA record, cited through ENISA's EUVD page. | |
| F11 editorial-advisory | · | (advisory, low confidence) The report's UNCTAD row omitted the Howard-Jones caveat and the Stamos quote added after iteration 1. | Applied: the row names both. | |
| F11 editorial-advisory | · | (advisory, low confidence) The 2026-08-06 section still points to the later ransomware use and Cadence breach '(2026-09-30 update)'. | Declined: a pointer to the later section, not a claim about the later events; the verifier says it may stay. |
Iteration #? CLEAN · 2 findings (truth=0, editorial=0, advisory=2) · Claude Sonnet 5.5 · 9m 56s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F11 editorial-advisory | · | (advisory, low confidence) The report and the W1-E row said the VWBP entry quotes German without translation in its main text; the running text glosses each German quote, and the untranslated German i | Applied: the report, the tracker row and the W1-E item describe the evidence quotes. | |
| F11 editorial-advisory | · | (advisory, low confidence, pre-existing) The Triage line's benign-lookalike examples are uncited; the discriminator that follows is source-backed. | Carried to the W1-E list for wave 02 (text this fire did not change). |
Iteration #? NEEDS_FIXES · 4 findings (truth=3, editorial=0, advisory=1) · Claude Sonnet 5.5 · 17m 48s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F3 claim-not-supported | · | The 2026-10-20 section and the Triage line made JetBrains' VPN advice an interim recommendation; the follow-up gives only the temporary access restriction as the interim step and says internet-facing | Applied: both passages give the temporary restriction as JetBrains' strong recommendation and the VPN or extra layer as a separate measure to consider. | |
| F3 claim-not-supported | · | (low confidence) 'as the CVE's numbering authority' and the 9.1 were cited to NLnet Labs' advisory page, which carries neither. | Applied: the 9.1 and the numbering-authority role are cited to ENISA EUVD-2026-80291 (assigner NLnet Labs, baseScore 9.1), added to sources[]; the advisory page is cited for the rating and vector only; the 2026-10-11… | |
| F3 claim-not-supported | · | (low confidence) 'The revision replaced the upgrade-tools load-error indicator' was cited to Bishop Fox, whose post predates the revision; the section repeated it uncited. | Applied: both passages say Bishop Fox's route never trips the indicator Check Point then listed and that the revised sk1000171 no longer lists it, each cited; this fire's record summary says Check Point revised its… | |
| F11 editorial-advisory | · | (advisory, low confidence) An uncommitted check_run.py change (the confirmation messages' iteration label) was in neither the report nor the record. | Applied: the report's change table and the record's notes name it. |
Iteration #? NEEDS_FIXES cap-breach · 4 findings (truth=1, editorial=0, advisory=3) · Claude Sonnet 5.5 · 13m 03s
| F-code | Section | Item · URL/quote | Verifier summary | Remediation · outcome |
|---|---|---|---|---|
| F3 claim-not-supported | · | (low confidence) The 2026-10-04 section still said the old indicator was 'replaced in the 2026-10-13 revision of sk1000171', uncited inside a Bishop Fox citation that predates the revision. | Applied after the pass: the parenthetical reads 'as sk1000171 then listed it'. Not re-verified (early exit). | |
| F11 editorial-advisory | · | (advisory, low confidence) 'TSSHOCK attacks on weak zero-knowledge proofs' fits two of Verichains' three attacks. | Applied: 'attacks on threshold-signing proofs and their encoding'. | |
| F11 editorial-advisory | · | (advisory, low confidence) 'per-agent cost-cap thresholds' is the entry's gloss; Mandiant places the caps at the Service ID and project level. | Applied: 'cost-cap thresholds and bounded recursion limits at the service-identity and project level'. | |
| F11 editorial-advisory | · | (advisory, low confidence) This fire's record summary did not mention that the title gives 8.4 again. | Applied: the record summary says so. |
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
2026-10-20T0835Z-audit · audit · Opus 5.5 · 0 entries published
Verification & coverage notes
Cleanup wave 01 of the one-time store cleanup program (prompts/cleanup-coordinator.md), an operator-directed partial audit with a partial claim; it ran no coverage re-sweep, so it does not move the weekly audit's window anchor. Report: docs/audits/2026-10-20-cleanup-wave-01.md. Tracker: docs/cleanup/STATUS.md.
Scope. § C0 bootstrap (tracker and plan on main as 0f2be64), then wave 01: the record tool and the gate false positives (§ C7 items 1 and 2), the one open store warning, and the 16 entries the 2026-10-18 audit's truth passes T1 to T9 found with factual errors outside its week. Those passes read the store before the 2026-10-11 audit's records landed, so the repair workers re-checked every defect against the current text.
Outcome. 13 entries carry a record from this fire (9 corrections, 3 updates, 1 improvement); Kairos Libercourt, Bitget and Apple CoreGraphics needed none, the 2026-10-11 audit having fixed all their listed defects. Two legacy entries (THORChain, actions-cool) left the legacy queue, 408 to 406 pending. The SolarWinds repair worker's CVE-2026-28324 workaround was removed by the main agent: no citable page carries it (EUVD's record, both trust-center advisories and SolarWinds' secure-configuration page were read). Registry: THORChain incident summary rewritten, one part-of edge (actions-cool to campaign:mini-shai-hulud, Socket 2026-09-24). cves_seen: last_seen moved for CVE-2026-102489 and CVE-2026-63077.
Standalone merges during the wave (each its own branch, auto-merged): 6c75002 record tool, 4ca3e3b page-check false positives, bfded2c migration tools removed, f06331d backlog trim and work/ chaff, 11ff607 site leftovers, 7864387 doc maps, f037aa0 tool suites in the gate, f5df6fe source decisions (2026-10-11 recommendations 5 to 7), 8678b51 registry summaries (Eurail, BeaverTail, InvisibleFerret). The wave commit itself carries one tool change: the verification-confirmation messages of check_run.py name the iteration from the iteration: key.
Warnings. One acknowledgment added: the 2026-10-18 audit's duration_seconds=117266 spans its rate-limit pause, as its own notes say. check_run.py --all ends 0 warn, 0 fail.
Budget. 8 of 20 spawns (work/2026-10-20T0835Z-audit/budget.tsv): one Explore inventory for § C0, two repair workers and five verifier passes, about 2.93 M sub-agent tokens in all (repairs 0.61 M, verifier 2.13 M). No spawn was blocked or retried, and no budget cut was needed.
Verification. Five passes of cti-verification. Iteration 1 (all 152 changed claims) returned NEEDS_FIXES with two truth and two editorial findings, among them an F4 on the SolarWinds detection line; iteration 2 found one stale report cell (F4); iteration 3 returned CLEAN; the confirmation pass, iteration 4, read all 158 claims cold and refuted it with three citation-precision findings (TeamCity's VPN advice, Unbound's numbering-authority attribution, Check Point's "replaced" indicator); iteration 5 confirmed those fixes and found the same Check Point wording in a third place. That one truth finding was fixed after the pass and published on the early-exit rule (truth plus editorial at most 2, no F1 or F4), so the residual is 1. Iteration 5 also found that tools/claim_ledger.py lists no claims for sentences a fire edits inside earlier sections; the tracker carries it as W7-14.
← Operations dashboard · run-record contract: docs/pipeline.md