14 verified findings from 1 run · the settled record for this UTC day, in the classic brief order.
Criticality
Kind
Topic
Region
TL;DR · the day in one read
01Two years inside North Korean C2 infrastructure produces a victim count, an EU government confirmation, and a contractor with access to 30 companies. Researcher Vangelis Stykas disclosed at Black Hat USA on 2026-08-05 that nearly two years of maintained access to North Korean actors' servers let him identify 1,640 impacted organisations across 57 countries, 700 to 800 of them with intrusions he calls "really damaging". Digitaal Vlaanderen, part of the Flemish Government in Belgium, confirmed to WIRED that Belgium's Centre for Cybersecurity notified it on 2026-03-03, that the affected workstation was isolated and exposed credentials rotated, and that the incident is contained. The dominant access route is the fake-job-interview lure, and the multiplier is compromised external contractors — Stykas saw some holding access to up to 30 companies. →
02Unit 42 finds the npm worm scraping runner process memory for OIDC tokens and building real Sigstore attestations over a backdoored tarball. Unit 42's analysis of CHAINDROP, the Shai-Hulud npm worm wave covered here on 2026-08-06, adds two mechanics that break controls defenders currently rely on. An embedded Python helper opens /proc/<pid>/maps and /proc/<pid>/mem on the GitHub Actions Runner.Worker process and searches live memory for OIDC tokens and runner secrets, so scanning files and environment variables at rest does not see it. A second, single-target path trades a runner OIDC token at npm's trusted-publishing endpoint for a real publish credential, injects a typosquatted dependency without touching install scripts, and then signs the result through Fulcio and Rekor — producing provenance Unit 42 is explicit is not forged. →
03Switzerland's national authority asks critical-infrastructure operators to block outbound RPC providers as compromised Swiss sites rise. NCSC-CH (BACS) published an advisory on 2026-08-07 reporting a rising count of compromised Swiss websites serving fake CAPTCHAs that instruct visitors to paste and run a command, and names the WP2Shell WordPress chain (CVE-2026-63030 with CVE-2026-60137) as what Swiss site operators and hosting providers have been reporting as the entry point. The pasted command pulls its next stage from a public blockchain reached through RPC-provider web interfaces, typically ending in an infostealer such as Vidar. BACS asks companies and critical-infrastructure operators outside fintech to restrict outbound connections to RPC providers — a concrete egress-policy change, not awareness advice. →
04The KVM shadow MMU yields a second guest-to-host escape, this one in the recursive zap path, with a public exploitation chain. A second use-after-free in the KVM/x86 shadow MMU, Zapscape (CVE-2026-64561), was assigned on 2026-08-04 and carried to European constituents by Belgium's Centre for Cybersecurity on 2026-08-07 in a "Patch Immediately" advisory that covers it alongside Januscape (CVE-2026-53359), the 2010-vintage bug this pipeline covered on 2026-07-09. Both are CVSS 8.8 and both let a root user inside a guest run commands on the host. Zapscape lives in the recursive zap path that runs during MMU page-quota reclaim, needs nested virtualization, and on Intel additionally requires EPT page-walk lengths 4 and 5 exposed to L1 — on AMD there is no such constraint. CCB notes RHEL-class distributions can let an unprivileged guest user reach guest root in the first place. →
05Kemp LoadMaster's pre-auth command injection moves from observed-but-failing attempts to a KEV listing, five weeks on. CISA added CVE-2026-8037 to its Known Exploited Vulnerabilities catalog on 2026-08-07, based on evidence of active exploitation of the unauthenticated command-injection flaw in Progress Kemp LoadMaster. When this pipeline last covered it on 2026-07-02 the only observed activity was exploitation attempts that eSentire reported as unsuccessful. Every LoadMaster running a version at or below GA 7.2.63.1, or the LTSF release 7.2.54.17, with the API enabled is affected; any appliance that sat internet-reachable and unpatched between the 29 June proof-of-concept and now warrants a compromise assessment rather than an upgrade alone. →
The interesting number in this disclosure is not the victim count. Researcher Vangelis Stykas told Black Hat USA on 2026-08-05 that nearly two years of maintained access to North Korean actors' servers — in some cases reaching the operators' own infected workstations — let him identify "1,640 companies across 57 countries" affected by the country's operations, with around 700 to 800 of them suffering intrusions he describes as "really damaging": company access, root access to servers, root access to AWS (WIRED, 2026-08-05).
The number that changes a defender's model is 30. For many of the impacted organisations, Stykas says, compromised external contractors — who often held developer keys or access to multiple systems — vastly increased the blast radius of a single successful attack: "I have seen a couple of contractors that had access to up to 30 companies" (WIRED, 2026-08-05). The initial access is the long-documented one: developers lured with fake job offers at high salaries and asked to download a program as a coding test, which silently installs malware — the technique Microsoft tracks as the Contagious Interview campaign, running since as early as 2022 (WIRED, 2026-08-05).
One European government body is named and has confirmed. A spokesperson for the Flemish government told WIRED: "We can confirm that we were notified of this incident on March 3, 2026 by the Centre for Cybersecurity Belgium (CCB), following the researcher's disclosure," adding that the affected workstation was isolated, potentially exposed credentials and access were revoked and rotated, and that on its investigation the incident has been contained and remediated (WIRED, 2026-08-05). Japan's CERT says it confirmed the researcher's findings and worked with AEON Smart Technology on remediation. Boston Children's Hospital, also named, says the incident involved a former contractor's personal device rather than its own systems (DataBreaches.net, 2026-08-07). Other named organisations did not respond to WIRED. Stykas attributes broadly to North Korean operations and names no tracked cluster.
Triage: this campaign's initial access looks like ordinary developer behaviour by design — an engineer running an unfamiliar project as part of a hiring process is a developer running an unfamiliar project. The discriminator available in telemetry is not the execution itself but its provenance and timing: a build or interpreter chain originating from a freshly cloned repository or a downloaded archive that no ticket, project or repository in the organisation accounts for, on an endpoint belonging to someone who is not being onboarded to that work.
We can confirm that we were notified of this incident on March 3, 2026 by the Centre for Cybersecurity Belgium (CCB), following the researcher’s disclosure,
WIRED (spokesperson for the Flemish government)
I have seen a couple of contractors that had access to up to 30 companies,
Beacon, a CRM platform built for the UK voluntary sector and holding data for around 1,500 organisations, published an incident update on 2026-08-04 that is more candid than most and worse than its early framing suggested. Its investigation "has confirmed that copies of database backups were made and likely downloaded by the unauthorised third-party", supported by evidence of a spike in activity during the incident timeline symptomatic of data leaving its systems; because it judges it highly unlikely to establish which data related to whom, its advice to customers is that "you may want to assume that all data that you store in Beacon, including attachment files, has been downloaded" (Beacon CRM, 2026-08-04).
The access path is the transferable part. Infosecurity Magazine reports that "Beacon revealed in its public statement that a compromised access key was used to gain access to its systems", with no detail published on how the key was obtained, and quotes the provider's characterisation that "This was more sophisticated than a simple compromised username and password" (Infosecurity Magazine, 2026-08-07). A programmatic key is not an account: it does not sit behind multi-factor authentication, does not trip impossible-travel logic, and in a multi-tenant platform it is frequently scoped to the platform rather than to a tenant — which is how a single credential becomes every customer's backup.
Encryption at rest did not close the gap either. Beacon states that while it stores data in an encrypted state, its experts have advised that on the available evidence it is possible the responsible party would have been able to decrypt it before copying it out (Beacon CRM, 2026-08-04). That is the expected outcome when the attacker holds an application-layer credential: the platform decrypts for its own legitimate operations, so a stolen key inherits that ability.
Downstream, individual charities are confirming and notifying separately. Victim Support published its own statement saying the evidence suggests "copies of database back-ups were made and likely downloaded by an unauthorised third party" and that it has reported the incident to the Information Commissioner's Office and the Charity Commission (Victim Support, 2026-08-04). Infosecurity names Myton Hospices, Sheffield Hospital Charity, Priscilla Bacon Hospice Charity and Rowcroft Hospice in the healthcare sector, plus homelessness charity The Clock Tower Sanctuary, with affected data including names, email addresses, telephone numbers and donation records (Infosecurity Magazine, 2026-08-07). No actor has claimed the breach and no leak-site listing has appeared.
Triage: an access-key compromise on a supplier platform produces no telemetry on the customer side at all — that is the defining property, and it is why the detection burden sits with the provider's own audit logging of key usage and egress volume rather than with anything a downstream charity could have seen. Where you operate the platform, the discriminator is the shape of the access, not its credentials: a valid key performing bulk reads or backup retrieval at a volume and hour outside its established pattern, against tenants it has never touched before.
our investigation has confirmed that copies of database backups were made and likely downloaded by the unauthorised third-party
you may want to assume that all data that you store in Beacon, including attachment files, has been downloaded
LevelBlue's SpiderLabs OpsCTI team is tracking a ScreenConnect distribution campaign whose lure has moved past the static fake-update page. Instead, it "recreates convincing software update and installation alerts by impersonating the Microsoft Store and Apple App Store while reproducing the look and behavior of trusted applications through dynamic modal dialogs and other interactive web elements", also imitating the Google Meet pre-join screen, with progress bars and camera and microphone permission prompts that behave the way the real dialogs do (LevelBlue SpiderLabs, 2026-08-07).
The installation chain is a batch script into PowerShell into an MSI installed silently through msiexec.exe /quiet with UAC elevation. The operational detail worth carrying is what happens next: "Because each installer is cryptographically bound to its corresponding relay server through the embedded public key, it automatically registers with the attacker's ScreenConnect instance once installed" (LevelBlue SpiderLabs, 2026-08-07). The agent is deployed at guest-level permission rather than full administrative rights, which keeps its footprint small and its behaviour closer to a legitimate support install.
The delivery infrastructure is built to survive the loss of any single component: thousands of near-identical phishing-framework deployments, payloads hosted on legitimate cloud object storage (AWS S3, Cloudflare R2), anti-automation gating through honeypot form fields, artificial delays and User-Agent filtering to Windows desktop clients only, and full victim fingerprinting — address, geolocation, ISP, browser, timezone, screen resolution — before an installer is served at all. Successful infections notify the operator in real time through the Telegram Bot API. LevelBlue assesses the supporting scripts as AI-assisted on the basis of unusually verbose documentation-style inline comments and emoji markers (LevelBlue SpiderLabs, 2026-08-07).
Triage: the benign lookalike is a real support session, and it is common. The discriminators the cited mechanics support, in order of strength: the relay hostname the client registers to; installation at guest-level permission with no corresponding helpdesk ticket; and the process lineage — a browser spawning a batch script or PowerShell that calls msiexec.exe /quiet, which is not how a user or an administrator installs remote-support software deliberately. The fingerprinting gate also means an analyst re-visiting the lure URL from a sandbox or a non-Windows client will usually be served a decoy rather than the installer, so failure to reproduce the payload is not evidence the report is wrong.
impersonating the Microsoft Store and Apple App Store
Because each installer is cryptographically bound to its corresponding relay server through the embedded public key, it automatically registers with the attacker's ScreenConnect instance once installed.
CISA published ICS advisory ICSA-26-219-01 on 2026-08-07 covering five vulnerabilities in Controller-Pilot Data Link Communications as implemented over the ATN-B1 standard — the data link that carries text clearances and instructions between air traffic controllers and flight crews worldwide, under Advisory Circular 90-117. The advisory's product version is vers:all/*, which is the honest way of saying this is a property of the standard rather than a defect in any implementation: "ATN-B1 CPDLC relies on legacy clear text unauthenticated radio frequency links" (CISA, 2026-08-07).
The five split into two effects. CVE-2025-71409 (CWE-306, CVSS 3.1 7.1) is the absence of authentication for VHF Data Link messages, which lets a rogue ground station inject CPDLC messages producing unexpected or misleading clearances; CVE-2025-71412 (CWE-754, 7.1) covers injection of false emergency or status messages, which CISA describes as potentially leading to misallocation of resources, operational confusion and improper responses by flight crews, controllers and ground operations. The remaining three are availability effects at CVSS 5.3: CVE-2025-71410 (Unnumbered Disconnect and malformed link-control frames terminating sessions and forcing reversion to voice), CVE-2025-71411 (broadcast control frames disconnecting multiple aircraft simultaneously, leading to controller overload) and CVE-2025-71413 (malformed or out-of-sequence frames at the X.25 layer causing repeated resets). Every one is carried out remotely over radio frequency (CISA, 2026-08-07).
Two statements from CISA bound this correctly, and both should travel with any onward summary. On consequence: the vulnerabilities "do not constitute an unsafe aircraft condition but can degrade operational safety margins by increasing workload, delaying safety-critical instructions, and reducing situational awareness". On likelihood, from the advisory's machine-readable CSAF record: they "are exploitable in a lab environment. However, they require very specific conditions to be met and are unlikely to be exploited outside of a lab setting" (CISA, 2026-08-07). The same record gives the remediation category as none-available for all five CVEs. There is no fix to schedule and no configuration to change.
There is a home-region thread: the advisory credits the report to "Martin Strohmeier of Armasuisse", the Swiss federal armaments enterprise (CISA, 2026-08-07).
This is carried for situational awareness in the transport sector rather than as an action item, and it is deliberately shipped without one. Nothing in an enterprise security stack touches an RF data link — the exposure belongs to air navigation service providers, airlines and aviation regulators, at the level of contingency planning for reversion to voice communication and of the multi-year standards work that would add authentication to the protocol. For a defender reading this brief, the useful takeaway is calibration: when reporting on this advisory circulates in less careful form, the two CISA statements above are what keep it in proportion.
ATN-B1 CPDLC relies on legacy clear text unauthenticated radio frequency links.
These vulnerabilities do not constitute an unsafe aircraft condition but can degrade operational safety margins by increasing workload, delaying safety-critical instructions, and reducing situational awareness.
These vulnerabilities in the CPDLC protocol stack are exploitable in a lab environment. However, they require very specific conditions to be met and are unlikely to be exploited outside of a lab setting.
CISA (CSAF record for ICSA-26-219-01)
vulnerability08 Aug 05:25Zsingle-source · national CERTOpen finding ↗
Apple's 2026-08-06 updates for macOS Tahoe 26.6.1, Sequoia 15.7.9 and Sonoma 14.8.9 fix CVE-2026-65400 in Screen Sharing, the VNC-based remote-desktop service built into macOS. Apple's description of the impact is unusually direct for a first-party bulletin: "An attacker on the network may be able to authenticate to Screen Sharing without valid credentials", with the cause given as "An authentication issue was addressed with improved state management" (Apple, 2026-08-06). NCSC-NL carried it to European constituents the following day, classing it CWE-287 (NCSC-NL, 2026-08-07). Apple publishes no severity score and reports no exploitation.
Taken alone this is a straightforward patch item on a client operating system. What raises it is the interval. One week earlier, macOS reverse-engineer fG! published an account of a separate defect in the same screensharingd daemon: a pre-authentication bug that lets an unauthenticated caller "download any file from a vulnerable macOS machine" given its full path, with /etc/sudoers offered as the worked example. He states "I know the bug was fixed by the DoS entry" in the preceding Apple bulletin, and argues that entry understates what the bug actually was (fG!, 2026-07-29). He is explicit that this is not the bug the other research team reported. That characterisation is his and Apple has not endorsed it, so it is context rather than a finding.
The operational consequence does not depend on adjudicating that dispute. Two independent, severe authentication and state-handling defects have now surfaced in the same daemon across consecutive release cycles, one of them described publicly in enough detail to be actionable. An estate that took 26.6 and deferred 26.6.1, or that reads bulletin severity labels as a patch-prioritisation input, ends up with the wrong picture in both directions.
Detection concept and hardening are the same lever: Screen Sharing listens on TCP/5900 and is discoverable over mDNS, so the exposure to hunt for is any managed Mac with the service reachable from a segment it does not need to serve — inventory it from network telemetry rather than from configuration policy, because the setting is per-device and users enable it. There is no known persistent artifact specific to CVE-2026-65400: Apple frames it as an authentication-state defect, not a post-exploitation primitive, so a successful attack looks like a legitimate Screen Sharing session, and the discriminator available is the session's source rather than anything about the session itself. Where the service is genuinely required, restrict it to a defined set of management sources or a VPN path. On the related daemon bug fG! is precise about the limits of the platform's own controls, and the limits cut both ways: he records that it "doesn't care about TCC either", but also that "It would be perfect if it could bypass SIP. That one it doesn't do" (fG!, 2026-07-29) — so System Integrity Protection does constrain that particular chain, and the file-download primitive it describes is a read, not a write.
An attacker on the network may be able to authenticate to Screen Sharing without valid credentials
An authentication issue was addressed with improved state management.
Flowise, the open-source visual builder for LLM and AI-agent workflows, picked up three CVEs on 2026-08-06 — and all three CVE records list the vendor's own sunset announcement (FlowiseAI) among their advisory references, which is the detail that turns a routine batch into an architecture decision.
The one that matters most needs no account. Per the assigning CNA, "Flowise through 3.1.4 contains an authentication bypass vulnerability that allows unauthenticated attackers to access the OAuth2 credential refresh endpoint by exploiting prefix-based whitelist matching in the authentication middleware defined in packages/server/src/utils/constants.ts" — a POST to the OAuth2 credential-refresh route with a trailing credential identifier appended slips past a check that only compares path prefixes, triggering unauthorised OAuth token rotation against credentials belonging to any workspace and potentially breaking dependent integrations (VulnCheck, 2026-08-07). CVE-2026-70636 is scored CVSS 4.0 8.7 with integrity-only impact (VC:N/VI:H/VA:N) and, notably, "This is a bypass of CVE-2026-41273" — the same route has been fixed once already and the fix was incomplete (VulnCheck, 2026-08-07).
The other two need an account but cross a tenancy boundary. CVE-2026-67622 (CWE-639, CVSS 4.0 8.5) is an insecure direct object reference in the OpenAI Assistants integration: an authenticated attacker supplies an arbitrary credential UUID to Assistants endpoints, the credential lookup performs no workspace-ownership check, and the attacker can enumerate cross-workspace assistant metadata, retrieve file and vector-store listings, and upload files into a victim workspace (VulnCheck, 2026-08-07). CVE-2026-67621 (CWE-862, CVSS 4.0 7.2) lets a member holding only view-level permissions call the document-store upsert and refresh routes directly to trigger ingestion, refresh vector-database contents, consume embedding API credits and modify knowledge bases that downstream chatflows depend on (VulnCheck, 2026-08-07). Germany's BSI CERT-Bund carried all three on 2026-08-06 with the summary that an attacker can exploit multiple Flowise vulnerabilities to bypass security measures, disclose information and manipulate data, marking the advisory unpatched with no fixed release listed (BSI CERT-Bund, 2026-08-06).
No party reports exploitation of any of the three. What makes this more than a routine batch is that the usual next step does not exist: with the company winding down commercial operations, an operator waiting for a fixed release is waiting for something nobody has committed to ship, and the code remaining available for community forks is not the same thing as a maintained security response. Self-hosted AI-agent orchestration platforms have been a recurring source of pre-authentication paths — this pipeline has covered three separate confirmed-exploited Langflow flaws since mid-July — and Flowise now belongs to the subset of that class where the only remaining controls are ones the operator builds.
Detection concept: for CVE-2026-70636 the network-visible shape is an unauthenticated POST to an OAuth2 credential-refresh path carrying an extra trailing path segment beyond the route the allow-list was written for, with the resulting token rotation appearing in the OAuth provider's audit log as a refresh nobody initiated. For the two authenticated flaws, the tell is a session enumerating credential UUIDs or Assistants endpoints outside its own workspace, and a view-only account issuing document-store upsert or refresh calls at all. Hardening, in the absence of a patch: terminate the credential-refresh route at a reverse proxy that enforces authentication independently of the application, scope each workspace's provider credentials so a cross-workspace read yields keys that are separately revocable, and treat any internet-exposed Flowise instance as a candidate for removal from the perimeter rather than for patching.
Flowise through 3.1.4 contains an authentication bypass vulnerability that allows unauthenticated attackers to access the OAuth2 credential refresh endpoint by exploiting prefix-based whitelist matching in the authentication middleware defined in packages/server/src/utils/constants.ts.
Ein Angreifer kann mehrere Schwachstellen in Flowise ausnutzen, um Sicherheitsvorkehrungen zu umgehen, Informationen offenzulegen und Daten zu manipulieren.
Cisco's IOS XE engineering team published a security hardening release on 2026-08-05 carrying seven CVEs, and the disclosure model matters more than any individual identifier. Rather than one CVE per bug, Cisco "grouped these issues by their underlying vulnerability class — Common Weakness Enumeration (CWE) — and assigned a single Common Vulnerabilities and Exposures Identifier (CVE ID) to each CWE grouping", stating plainly that "the CVSS score that is assigned to each CVE ID represents the maximum potential severity of the single most impactful underlying bug within that specific CWE category" (Cisco PSIRT, 2026-08-05).
The consequence for anyone running a risk-based patch process is that per-flaw triage is unavailable by construction. CVE-2026-20272 carries CVSS 9.8 for improper neutralisation of special elements — command, OS and argument injection — but that score belongs to the worst bug in the group, and neither the count of bugs behind it nor their individual reachability is published. The remaining six are CVE-2026-20267 (improper access control, 9.0), CVE-2026-20268 (memory-buffer bounds, 8.6), CVE-2026-20269 (resource lifetime, 8.6), CVE-2026-20270 (incorrect calculation, 8.6), CVE-2026-20271 (control-flow management, 8.6) and CVE-2026-20273 (input validation including path traversal, 8.6) (Cisco PSIRT, 2026-08-05). The advisory's aggregate CVSS vector is AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H — network-reachable and unauthenticated at the top of the range.
Exposure is broad and configuration-independent: the vulnerabilities "affect Cisco IOS XE Software when it is running in autonomous or controller mode, regardless of device configuration", across releases 17.9, 17.12, 17.15, 17.18 and 26.1, with first fixed releases 17.9.10, 17.12.8, 17.15.6, 17.18.4 or 17.18.4a, and 26.1.2 respectively; Catalyst 3650 and 3850 Series switches run none of these trains and were not evaluated (Cisco PSIRT, 2026-08-05). Cisco states "there are no workarounds that address these vulnerabilities" and that the flaws "were found during internal testing and are not known to be actively exploited", with PSIRT aware of no public announcements or malicious use (Cisco PSIRT, 2026-08-05). NCSC-NL relayed the release to European constituents on 2026-08-07 (NCSC-NL, 2026-08-07).
One line in the Source section is worth reading twice: "These vulnerabilities were found during internal security testing using existing testing processes as well as frontier AI models" (Cisco PSIRT, 2026-08-05). A vendor attributing a bulk hardening release partly to model-assisted review is a plausible signal that such releases become more frequent and larger, which is a planning input for change windows on network infrastructure rather than an immediate threat.
No detection guidance is possible here and none should be attempted: Cisco publishes no per-bug technical detail, no reachable component, and no exploitation pattern, so nothing supports a hunt hypothesis. The advisory does list Snort rules 66897-66898 as associated coverage. This is an inventory-and-schedule item — identify every device on 17.9, 17.12, 17.15, 17.18 or 26.1, and move it to the first fixed release for its train — carried here because there is no configuration that removes the exposure and no interim mitigation to fall back on, not because anything indicates it is being attacked.
The CVSS score that is assigned to each CVE ID represents the maximum potential severity of the single most impactful underlying bug within that specific CWE category.
These vulnerabilities were found during internal testing and are not known to be actively exploited.
There are no workarounds that address these vulnerabilities.
These vulnerabilities were found during internal security testing using existing testing processes as well as frontier AI models.
Wiz Research's semi-annual cloud threat report covers January to June 2026, and its value for this constituency is the named inventory rather than the trend lines: it says concretely which AI infrastructure attracted attacker and researcher attention, and what the resulting exposure looks like in a cloud estate.
The AI toolchain now has its own vulnerability cadence. LiteLLM — an AI gateway Wiz says is present in over a third of the cloud environments it monitors — "had four separate security events in six months: a supply-chain compromise, an SQL injection vulnerability exploited in the wild, a privilege escalation chain and an authentication bypass", while Dify, Langflow, n8n and Ollama "each had critical unauthenticated vulnerabilities of their own" (Wiz Research, 2026-08-06). That list is worth reading as an asset-inventory prompt: these are components teams stand up quickly, often outside the change process that governs the rest of the estate, and three of the five have already reached this pipeline's coverage through separate exploited-vulnerability events.
The exposure finding is sharper than the vulnerability one. On Model Context Protocol servers, Wiz reports: "We found unauthenticated MCP endpoints across hundreds of environments, each one a pre-authenticated proxy holding backend credentials and bridging multiple services" (Wiz Research, 2026-08-06). The reason that shape matters is that an MCP server is not a data store to be broken into — it is a component that already holds the credentials for everything behind it and exists to act on their behalf, so reaching it unauthenticated is not a step toward access, it is the access.
On the actor side, Wiz profiles JINX-0163, a cloud-native extortion group it began tracking in 2026 and that "consistently targets non-human identities - service accounts and IAM roles - rather than end users", in some cases leveraging a single over-privileged identity or an exposed state file to pivot to a full inventory (Wiz Research, 2026-08-06). An extortion group that skips human identity entirely bypasses most of the control stack organisations have spent two years building — phishing-resistant MFA, conditional access, helpdesk verification — none of which applies to a service account.
On supply chain, Wiz records that notable supply-chain attacks "went from making up about 10% of significant incidents in H2 2025 to 25% in H1 2026", with TeamPCP, North Korea and at least three independent operations running campaigns concurrently across npm, PyPI, Composer, VSCode extensions, Jenkins plugins and AUR, several of which had not been targeted this way before (Wiz Research, 2026-08-06). It also notes that malicious packages' shrinking availability window is what makes an install cooldown policy effective — declining to download packages published less than 24 hours ago — which is a specific, cheap control rather than a general recommendation.
We found unauthenticated MCP endpoints across hundreds of environments, each one a pre-authenticated proxy holding backend credentials and bridging multiple services.
They went from making up about 10% of significant incidents in H2 2025 to 25% in H1 2026.
Elastic Security Labs published endpoint telemetry from a macOS host in which shells running under Claude Code scripted a login to an ephemeral tunnel hostname, pulled application metrics, "stood up a Cloudflare quick tunnel, and installed LaunchAgent persistence", leaving a local application reachable from the internet across reboots (Elastic Security Labs, 2026-08-07).
The value is that this is observed rather than constructed. The technique class — an AI coding agent steered into doing something the developer did not intend — has been demonstrated in laboratory conditions before; what Elastic contributes is what it looks like in production telemetry, and the answer is that it looks like work. "Coding agents such as Claude Code and Cursor are vendor-signed, used all day on developer laptops, and routinely open shells, call APIs, edit files, and install helpers" (Elastic Security Labs, 2026-08-07). Every heuristic that normally carries weight — code signature, process reputation, whether the parent is a known-good binary, whether shell invocation is expected from this tree — returns the reassuring answer. Elastic also notes the immediate children were often shells (zsh) and helpers under that ancestry rather than the agent executing every binary itself, so lineage depth matters when writing the logic. Alongside that full chain it reviewed shorter cases on other hosts with a coding agent still the execution parent — Claude Code staging JavaScript under /tmp through Apple-signed Python and osascript; a Cursor session that attempted a decrypted keychain dump filtered toward Linear and Model Context Protocol OAuth material, which endpoint controls blocked; and Claude Code with permission bypass pulling an unsigned binary over plaintext HTTP, attempting quarantine stripping and ad-hoc re-signing.
Elastic declines to call it malicious, and makes that the point rather than a hedge: agent-parented reverse tunnels and LaunchAgents can expose a local admin application to the internet, and its guidance is to "Treat that as high severity even when it looks like vibe-coded ops, not confirmed malware" (Elastic Security Labs, 2026-08-07). That is the right severity model for this class. Whether the agent was steered by an attacker or simply took an over-broad route to a task the developer asked for, the resulting exposure is identical, and waiting to establish intent before acting means waiting past the point where the tunnel is already up.
Triage: developer endpoints legitimately produce every one of these events in isolation — tunnels for previewing local work, LaunchAgents for local services, agents spawning shells constantly. The discriminators Elastic's case supplies are the conjunction and the durability: an outbound tunnel plus a persistence mechanism that survives reboot, both under the same agent ancestry, is not a shape that ordinary preview-and-iterate work produces, because a preview tunnel has no reason to outlive the session.
Coding agents such as Claude Code and Cursor are vendor-signed, used all day on developer laptops, and routinely open shells, call APIs, edit files, and install helpers.
stood up a Cloudflare quick tunnel, and installed LaunchAgent persistence
Treat that as high severity even when it looks like vibe-coded ops, not confirmed malware.
The interesting part of Check Point Research's Black Hat disclosure is where the bugs are, not how many there are. All five sit in workerd's own native code — four of them memory-corruption defects in the "glue" layer — the C++ code that marshals data between JavaScript and native implementations — which is the seam that every isolate-based multi-tenant runtime depends on and that JavaScript-level reasoning about sandbox safety does not cover (Check Point Research, 2026-08-06).
Three are worth naming for what they say about the class. An out-of-bounds read in the URLPattern implementation arises from a mismatch between the capture-group count workerd's own parser computes and the count V8's regex engine actually produces. Two use-after-frees come from native-object lifetime management: one in node:zlib's deflateParams(), one in HTMLRewriter's AttributesIterator. The fifth is not a memory-corruption bug at all: a SQL authorization bypass in the Durable Objects storage path that Check Point calls "a classic that leads to arbitrary deserialization" (Check Point Research, 2026-08-06).
Two chains were demonstrated, and the second is the reason this belongs in an operational brief rather than a conference recap. The first is a cross-tenant heap read: one Worker reaching across the shared process heap to read a co-located tenant's secrets. The second starts from a prompt injection into Code Mode — Cloudflare's LLM tool-use feature — and rides the zlib use-after-free out of the V8 isolate into native code execution on the host. Check Point's own framing is that "Because workerd underpins both Code Mode sandboxes and Workers tenant isolation, the findings create sandbox-escape and cross-tenant exposure risk" (Check Point Research, 2026-08-06).
That chain is a concrete instance of something the AI-security discussion usually leaves abstract. Prompt injection is generally reasoned about as a content problem — the model can be made to say or request the wrong thing — with the sandbox as the backstop that bounds the damage. Here the model-controlled code is the input that reaches a memory-corruption bug in the sandbox itself, so the backstop is what fails. Exploitation still requires getting attacker-chosen JavaScript to run inside a Worker, which in the managed platform means being a tenant, and in the Code Mode case means steering the model.
Remediation is uneven in a way that matters. Cloudflare's managed Workers environment has been fixed in production, but "Self-hosted workerd / Code Mode deployments should update to v1.20260619.1", and "As of now, Cloudflare has not assigned CVEs" (Check Point Research, 2026-08-06). Check Point released proof-of-concept code as part of the presentation. For most readers the managed fix means no action; for anyone running workerd themselves the absence of a CVE means no scanner, SBOM tool or advisory feed will surface this — the version check has to be made deliberately. No in-the-wild exploitation is reported.
Triage: no host-side detection concept follows from what is published — the exploitation is in-process inside a runtime that does not expose per-isolate telemetry to its operators, and Check Point describes no post-exploitation artifact. The honest operational content here is the version check and the design lesson, not a hunt.
Because workerd underpins both Code Mode sandboxes and Workers tenant isolation, the findings create sandbox-escape and cross-tenant exposure risk.
Self-hosted workerd / Code Mode deployments should update to v1.20260619.1.
Switzerland's national cyber authority has attached its own jurisdiction's numbers to the WordPress chain this pipeline recorded reaching CISA KEV in July. In an advisory published 2026-08-07, BACS reports a rising count of compromised websites presenting fake CAPTCHAs that push visitors into executing a command themselves, puts the worldwide population of compromised sites at more than 100,000, and states it is currently seeing an increase in the number of Swiss websites being compromised and used to distribute malware (NCSC-CH, 2026-08-07).
The entry point is the already-covered one, now with reporting behind it: BACS writes that in recent days it has received an accumulation of reports from Swiss website operators and web-hosting providers describing exploitation of two recently disclosed WordPress vulnerabilities, and that "criminals use a combination of two vulnerabilities in WordPress" known as WP2Shell — CVE-2026-63030 and CVE-2026-60137 (NCSC-CH, 2026-08-07). Most of the compromised sites run WordPress.
What is new below the entry point is the delivery chain. Once a visitor follows the instruction and runs the command, it fetches further malicious code whose storage and distribution sit on a public blockchain — the technique BACS names EtherHiding — retrieved through the web interfaces of RPC providers that broker access to those networks; the payload is typically an infostealer, with Vidar named as an example, going after credentials, payment-card data and cryptocurrency wallets (NCSC-CH, 2026-08-07). That hop is why the authority's recommendation is an egress-policy one rather than a filtering one: a blockchain read has no domain to sinkhole.
Triage: the client-side execution has a distinctive shape in process-creation telemetry with parent lineage — a command interpreter (powershell.exe on Windows, the terminal shell on macOS) started from a browser process tree, immediately followed by outbound HTTP to an RPC-provider endpoint. Neither half is individually rare on a developer or administrator workstation; the sequence, on a general-office endpoint, is the signal, and the browser parentage is what separates it from legitimate admin scripting, which is not launched from a browser. On the server side, the compromise signature is the WP2Shell request pattern against the unauthenticated REST batch endpoint in web-server access logs, followed by administrator-account or plugin and theme file changes that no admin action accounts for.
Dabei verwenden Cyberkriminelle eine Kombination aus zwei Schwachstellen in WordPress
Unternehmen und Betreiberinnen kritischer Infrastrukturen, welche nicht im Fintech-Bereich tätig sind, sollten ausgehende Verbindungen zu RPC-Anbietern einschränken.
Januscape has a sibling. Zapscape (CVE-2026-64561) was assigned on 2026-08-04 and is a use-after-free in the same KVM/x86 shadow MMU, but in a different path: where Januscape came from shadow-page role confusion, Zapscape fires when a guest using nested virtualization makes KVM recursively zap a root shadow page that is still in use during MMU page-quota reclaim, leaving KVM handling a fault on a root that has already been invalidated (V4bel, 2026-08-06). The upstream fix, commit 2abd5287f083, moves the stale-root check after make_mmu_pages_available() (V4bel, 2026-08-06).
Belgium's Centre for Cybersecurity turned the pair into a single operational instruction on 2026-08-07, publishing a "Patch Immediately" advisory that treats both CVEs together, scores each CVSS 8.8 (AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H), and states that they "can let an attacker run commands on the host by escaping the guest virtual machine, this can lead to a full system compromise since the attacker can escape the virtual environment" (CCB, 2026-08-07).
Two preconditions decide who is actually exposed, and they differ between the two bugs. Januscape "can be triggered on Intel without any particular constraint as long as nested virtualization is enabled, but Zapscape requires that both EPT page walk length 4 and 5 be exposed to L1 on Intel. On AMD there is no such constraint" (V4bel, 2026-08-06); CCB states Zapscape "can only be exploited by attackers with root privileges and systems with Intel architecture need both EPT page walk length 4 and 5 exposed to L1" (CCB, 2026-08-07). The guest-root precondition sounds like a high bar and often is not: CCB adds that "Distributions like RHEL can allow an unprivileged user to gain root privileges," and on rented cloud instances guest root is simply what the tenant has (CCB, 2026-08-07). The researcher's current public demonstration runs on AMD nested SVM/NPT (V4bel, 2026-08-06).
These two vulnerabilities can let an attacker run commands on the host by escaping the guest virtual machine, this can lead to a full system compromise since the attacker can escape the virtual environment.
This vulnerability can only be exploited by attackers with root privileges and systems with Intel architecture need both EPT page walk length 4 and 5 exposed to L1.
Distributions like RHEL can allow an unprivileged user to gain root privileges.
can be triggered on Intel without any particular constraint as long as nested virtualization is enabled, but Zapscape requires that both EPT page walk length 4 and 5 be exposed to L1 on Intel. On AMD there is no such constraint.
Unit 42 published its own analysis of the CHAINDROP wave on 2026-08-06, and two of its findings change what defenders can rely on rather than adding detail to what they already knew.
The first is credential theft that never touches disk. An embedded Python helper hidden inside an encrypted blob in the payload "locates the Runner.Worker process on GitHub Actions runners, opens /proc/<pid>/maps and /proc/<pid>/mem, and searches live process memory for OpenID Connect (OIDC) tokens and runner secrets" (Unit 42, 2026-08-06). Ephemeral OIDC tokens exist to avoid long-lived secrets sitting in a file or a variable; reading them out of the runner's address space while they are live defeats that design, and any secret-scanning control that inspects files or environment variables at rest sees nothing.
The second is a single-target path that is worse than a forgery. The worm checks three environment variables and only proceeds if it finds itself inside GitHub Actions, in a repository whose name contains /opensearch-js, in a workflow whose reference contains release-drafter.yml; anywhere else in that project it exits and steals nothing, staying silent in exactly the runs a maintainer is most likely to be reading (Unit 42, 2026-08-06). In that path it asks the runner for an OIDC token scoped to npm:registry.npmjs.org and trades it at npm's own trusted-publishing exchange for a real publish credential — the repository's legitimate release identity becomes the attacker's. It then downloads the latest tarball, bumps the patch version and adds a single dependency line typosquatting the project's own scope, never touching install scripts at all, so detections built around preinstall hooks would miss it. Finally it requests a second OIDC token for Sigstore, obtains a Fulcio certificate, builds an in-toto SLSA v1 provenance statement over the tarball's SHA-512 hash, signs it and uploads the entry to the public Rekor transparency log (Unit 42, 2026-08-06).
Unit 42 is explicit about what that means: "This is not forged provenance. The attestation says the tarball was built in that repository by that workflow, and that is true." Its guidance follows directly — a package having valid npm provenance does not mean the package is clean, only that the tarball came out of the workflow named in the certificate, and if that workflow is running attacker code then valid provenance is what you should expect to see. "Pivot on the Rekor log index and the workflow identity inside the certificate, not on whether the signature checks out" (Unit 42, 2026-08-06).
Unit 42 states it did not observe this path execute and that it cannot execute anywhere except in that one workflow in that one repository, but that it is fully implemented and reachable from the payload's main entry point (Unit 42, 2026-08-06). The worm also runs a locale gate before any collection, exiting cleanly on a Russian-language host.
locates the Runner.Worker process on GitHub Actions runners, opens /proc/<pid>/maps and /proc/<pid>/mem, and searches live process memory for OpenID Connect (OIDC) tokens and runner secrets.
This is not forged provenance. The attestation says the tarball was built in that repository by that workflow, and that is true.
Pivot on the Rekor log index and the workflow identity inside the certificate, not on whether the signature checks out.
CISA added CVE-2026-8037 to its Known Exploited Vulnerabilities catalog on 2026-08-07, "based on evidence of active exploitation" (CISA, 2026-08-07). The catalog record describes the flaw as a command injection that "allows an un-authenticated attacker to execute arbitrary commands on the LoadMaster appliance by exploiting unsanitized input in multiple command endpoints," classes it CWE-77, and records known ransomware-campaign use as unknown (CISA, 2026-08-07).
The delta is the status, not the mechanics. This pipeline's 2026-07-02 entry recorded exploitation attempts beginning the day the proof-of-concept dropped, all of them unsuccessful with no post-compromise activity; a federal catalog entry asserting active exploitation is a different claim, arriving five weeks later. Nothing in the affected estate has changed: watchTowr Labs gives the vulnerable version range as "Kemp LoadMaster: GA v7.2.63.1 and older" together with the LTSF release v7.2.54.17 and older, in both cases only when the API is enabled (watchTowr Labs, 2026-06-29). No authority has named an exploiting cluster or described an observed intrusion path.
The catalog's remediation due date is a US federal compliance clock and carries no weight here. What does carry weight is the interval: a public exploit has existed since late June against an appliance class that terminates traffic at the network edge, and the flaw needs nothing but reachability to the API. An organisation that patched in June is fine. An organisation that has been treating this as a scheduled item now has a gap between the PoC and its own patch date during which a working, public exploit was being fired at exposed instances.
Detection remains network-side rather than host-side, because the appliance does not normally surface process telemetry to defenders: in reverse-proxy or web-application-firewall logs in front of the management API, unauthenticated POST requests to the /accessv2 endpoint carrying malformed or oversized parameters, and repeated probing of that endpoint from related sources in a short window, are the observable shape (watchTowr Labs, 2026-06-29). Triage: legitimate LoadMaster API clients authenticate and send well-formed payloads from a small, stable set of management sources — the discriminators are an unauthenticated request reaching /accessv2 at all, and parameter content that is malformed rather than merely unexpected. Hardening is unchanged and still the strongest control available: disable the LoadMaster API where it is not required, which removes the endpoint entirely, and keep the management interface off any general-purpose network.
Progress LoadMaster contains a command injection vulnerability that allows an un-authenticated attacker to execute arbitrary commands on the LoadMaster appliance by exploiting unsanitized input in multiple command endpoints.
Enumerate which external contractors and their personally-managed devices currently hold developer keys or standing access to your systems, and scope each one by what a compromise of that single device would reach — this campaign's blast radius came from contractors holding access to many organisations at once, not from breaching each organisation separately.
Restrict outbound connections from user endpoints to blockchain RPC-provider web interfaces, per BACS's own recommendation for companies and critical-infrastructure operators that are not in fintech — this is the takedown-resistant hop the pasted command depends on.
Confirm every WordPress instance the organisation or its agencies publish is on 6.9.5 / 7.0.2 or later, and for any site that was reachable and unpatched, check for attacker-added administrator accounts and modified theme or plugin files rather than only applying the update.
Patch KVM host kernels so they carry both fixes (Januscape commit 81ccda30b4e8 and Zapscape commit 2abd5287f083) — there is no guest-side mitigation for either, so a patched guest on an unpatched host is still exposed.
Turn nested virtualization off for every tenant and workload that does not explicitly need it; where it must stay on for Intel guests, restrict exposure of EPT page-walk lengths 4 and 5 to L1, which is Zapscape's stated precondition.
Enumerate every long-lived programmatic access key held by or issued to your SaaS suppliers and processors, and confirm each one is scoped below the level at which it could read a whole-tenant database backup — this breach turned one key into every customer's data, and no user credential or MFA control was in the path.
Stop treating a valid npm provenance attestation as evidence a package is clean; where release pipelines gate on provenance, change the check to pin on the workflow identity inside the Fulcio certificate and the Rekor log index, which is what Unit 42 says still discriminates.
Re-verify every Kemp LoadMaster is on GA 7.2.63.2 or the corresponding LTSF fixed build from Progress's June bulletin and, for any appliance that was internet-reachable with the API enabled before it was patched, run a compromise assessment of the appliance rather than closing the ticket on the version string.
Inventory Model Context Protocol endpoints in your cloud estate and confirm each one requires authentication — Wiz found unauthenticated MCP endpoints across hundreds of environments, and each is a pre-authenticated proxy already holding the backend credentials it bridges.
Alert on ScreenConnect client installation where the relay hostname is not one your organisation operates — the campaign's installers self-register to an attacker relay on install, so the relay identity is the discriminator that survives every cosmetic change to the lure.
If you run self-hosted workerd or Code Mode rather than Cloudflare's managed platform, update to v1.20260619.1 — there is no CVE to catch this in a scanner feed and no advisory will arrive through vulnerability-management tooling.
Turn Screen Sharing off on every managed Mac that does not actively need it — the flaw is reachable by an attacker on the network with no credentials, and the setting is the only control that does not depend on which of the last two macOS updates a device has taken.
Inventory self-hosted Flowise instances at ≤3.1.4 and put the OAuth2 credential-refresh route behind network or proxy-level authentication now, since no fixed release is coming — the unauthenticated path is reachable by appending a trailing credential identifier to an otherwise allow-listed route prefix.
2026-08-08T0409Z-intel· Claude Opus 5 · window 26 h · 14 entries published
Verification & coverage notes
Fourteen entries from twenty candidates. Four ship as delta updates on prior coverage rather than as new entries, two of them because the store-wide CVE index caught coverage older than the 14-day in-context window — a check that changed the disposition of both items after they had already been triaged as new.
Corrections applied before composition. Four claims returned by research did not survive verification against the authority that owns them, and all four would have reached the store:
A Flowise CVE was returned as "CVSS 9.9 critical". The assigning CNA's own per-CVE record scores CVE-2026-67622 at CVSS 4.0 8.5, and Germany's BSI publishes a single advisory-level score of 7.7 with no per-CVE breakdown. The authority governs; 9.9 appears nowhere in this run's output. All three Flowise records were cross-checked against the CVE data mirrored on OSV before any score entered frontmatter.
The two KVM vulnerabilities were returned as "two 16-year-old bugs". Only Januscape traces to the 2010 commit; Zapscape is a separate defect in a different code path, assigned four days ago.
The aviation advisory was returned without the publisher's own likelihood assessment. The machine-readable record behind the advisory states the flaws require very specific conditions and are unlikely to be exploited outside a lab setting, and records the remediation status as none-available. That statement is now the entry's calibration and its priority follows from it.
The npm-worm delta was returned as the worm minting "forged" provenance attestations. The cited analysis says the opposite in as many words — the attestation is genuine and truthfully records which workflow built the tarball, which is precisely why it is worse than a forgery. The entry carries the source's framing.
Sound and complete. The completeness sweep re-read every returned item including those the sub-agents themselves flagged borderline. Two flagged-borderline items were promoted into the publish set on review (the aviation protocol advisory, at the lowest priority and with no action item, because the transport sector is inside the constituency and severity doubt on a clearly relevant item resolves toward inclusion; and the macOS Screen Sharing fix, because a network-reachable authentication bypass on a remote-desktop daemon carries a concrete non-patch control). One item the sub-agent did not flag was dropped on review. No item was dropped for space.
No deep dive this window, and none manufactured. The two candidates with the technical depth a deep dive needs — the exploitation confirmation on the load-balancer flaw and the hypervisor escape — are both delta updates that must carry only what is new, and a long-form treatment would have meant recapping coverage the reader already has. The highest-relevance new item rests on journalism rather than technical analysis, and there is no kill chain in the sources to map.
borderline-drop: Elastic — npm min-release-age removal invisible to log-tailing telemetry — a real detection-engineering point, but framed around one vendor's own agent integration, and the underlying principle (the removal of a control is an event that append-only log tailing structurally cannot see) is already familiar to this audience.
borderline-drop: SSD Disclosure — Linux kernel net/bridge STP timer use-after-free — a no-CVE local privilege-escalation primitive, already fixed upstream, that the ordinary kernel patch cycle handles; also outside the window.
borderline-drop: Bol / De Bijenkorf customer-data breach via CEVA Logistics — no disclosed intrusion mechanism, no actor, and no transferable technical lesson; the shared-supplier access-path lesson is carried with far more substance by the Beacon CRM entry in the same window.
borderline-drop: Levi Strauss regulatory filing tied to the UNC6671 extortion ecosystem — the filing names neither actor nor technique, and the campaign link reaches this run only as a third-hand relay; yesterday's entry on that actor already carries the mechanics, and the defender action is unchanged.
borderline-drop: press analysis narrowing the Swiss federal SharePoint intrusion to two candidate CVEs — a journalist's "potentially either" inference, not a confirmation by the affected agency or the national authority. Publishing it would bind unconfirmed identifiers to a home-region incident in the store's CVE index, where automated consumers would read them as established.
borderline-drop: the Ransom Cartel operator's 16-year sentence — a law-enforcement outcome with no change to what a team patches, hunts, blocks or detects. Better suited to the weekly's law-enforcement lens.
borderline-drop: three leak-site-only victim claims surfaced through a leak-site tracker, including one Swiss-listed and one German company — no victim statement, no regulatory filing and no high-reliability journalism corroborates any of them, which is the fake-news guard working as intended rather than a coverage gap. Flagged for a corroboration re-check next run, since a confirmed Swiss victim would be squarely in scope.
Recency carve-out: the North Korean victim-set disclosure has a primary dated 2026-08-05, inside the 72-hour developing window rather than the 26-hour window. It reached this run through an in-window pickup, the story is still producing named-victim statements, and a confirmed European government victim keeps it in scope. Recorded in the entry's own sourcing note so no reader is misled about freshness.
Single-source: seven entries ship without a second independent source, each with the situation named in its own sourcing note. Two take the national-authority carve-out (the Swiss advisory for its own jurisdiction; the aviation advisory, where the publishing authority is the coordinating discloser for a standards-level finding with no vendor). Five are research labs reporting their own original work — the npm-worm delta, the runtime memory-corruption research, the coding-agent telemetry, the remote-support-tool distribution campaign, and the semi-annual cloud report — where no second party observed the same thing and none is claimed to have.
Credibility ratings follow corroboration rather than publisher count. Where a vendor advisory reaches this run through a national-CERT relay, that is one assessor with two publishers and the rating stays at 2; only items where a second party independently observed or assessed the thing carry 1.
Coverage gaps: google-tag (recipe drift — configured path resolves to the general security blog with no dated TAG listing; the one item that mattered was reached through another record and was already covered); recordedfuture-insikt, trellix, infoguard-labs, paradigm-shift-research (documented recipe gaps, not re-attempted); claroty-team82 (listing carries no publication dates; the one article drilled resolved to June); prodaft (stale upstream cache, fourth run); msrc-blog, siemens-productcert-csaf, flatt-security (not drilled — deprioritised against higher-yield sources inside the time budget); cert-at, cert-pl, cert-eu, ncsc-ie, enisa, ncsc-uk, le-monde-info, synacktiv, truesec (all fetched cleanly with nothing inside the window).
Allocation defect, this run's own: the rotation-priority line sent to S1 named prodaft, but no prodaft record was in S1's slice, so S1 could not attempt it. S2 and S3 both carried the record and both attempted it, so no coverage was lost. Fixed by making the rotation-priority list per-domain rather than shared.
Essential-coverage: no misses — all fifteen essential-tier records were attempted across S1 and S2.
The pinned ATT&CK dataset is at v19.1 with v19.2 available upstream (published 2026-08-05). Not updated mid-run: a release can revoke identifiers that immutable published entries already carry, which would create store-wide warnings this run could not fix. Left to the weekly, which owns the pin check.