2026-10-15T0256Z-intel

Verification & coverage notes

Window: gap 24.0 h from 2026-10-14T0255Z-intel, window 26 h, developing-story window 72 h; the container clock agreed with the network date (skew 1 s). Four new entries (Cisco Secure Email Gateway CVE-2026-76519 critical; PAN-OS CVE-2026-0322 high; Collabora Built-in CODE Server CVE-2026-110447 high; the fake WatchGuard VPN client site notable) and two changelog records (an update on the PK Softech/Publica entry, an improvement on the Danfoss refrigeration-controller entry). No deep dive: nothing cleared the bar. The previous record's publish_status is ok.

KEV sweep (work/2026-10-15T0256Z-intel/kev-window.txt): four additions dated 2026-10-14; CVE-2026-76519 and CVE-2026-0322 were NOT COVERED and became entries, the AhsayCBS pair (CVE-2026-105133, CVE-2026-105134) was already covered with exploitation stated. No RANSOMWARE rows.

Backlog (state/coverage_backlog.md): the one open row, Beyond Gravity, is unchanged (S2 and S4 re-checked: swissinfo, Netzwoche and bluewin of 2026-10-02 to 06 are still the freshest, no vector, actor or data class named, no leak-site listing); it stays held on its stated condition and expires 2026-10-20. No row changed state, so nothing was appended.

borderline-drop lines:

• borderline-drop: Fortinet October PSIRT batch FG-IR-26-176 to -185 (S1; FortiSIEM 7.5.0 CVE-2026-106173 9.4, FortiSandbox GUI authentication bypass CVE-2026-107656 9.6); Known Exploited is No on all ten, the FortiSIEM flaw affects 7.5.0 only and restricting TCP 27918 prevents it, NCSC-NL and CERT-FR only relay it; the 2026-10-14 fire dropped the same set on the same facts, so this is the same call, not a new one.
• borderline-drop: Juniper Junos OS and Junos OS Evolved October bundle, 153 records (S1; three unauthenticated J-Web paths to root CVE-2026-101956, CVE-2026-101986 and CVE-2026-101954 and unauthenticated overflows at 9.8), a quarterly vendor release, Juniper SIRT and CISA's SSVC record state no exploitation, J-Web needs the web-management service configured, the advisory portal is a JavaScript shell on every transport so only the CNA record (not citable) and one ENISA page were readable, and the Cisco Nexus OAM and Fortinet bundles were judged the same way. Re-gate when an exploitation report, a KEV listing or a CERT or press page that can be cited appears.
• borderline-drop: Microsoft October Patch Tuesday residual (S1; five Outlook preview-pane RCEs CVE-2026-70242, CVE-2026-97749, CVE-2026-97016, CVE-2026-88156, CVE-2026-70291, about 60 further unauthenticated network flaws at CVSS 9.0 or above, the .NET WDAC bypass CVE-2026-108398, MessagePack-CSharp CVE-2026-92707), none exploited, all rated Exploitation Less Likely or Unlikely, no public proof of concept for the unauthenticated items; the 2026-10-14 fire dropped this set on identical facts.
• borderline-drop: Chrome 155 stable, 37 fixes with seven Critical (S1), no exploitation statement, source dated 2026-10-13 20:53Z before the window.
• borderline-drop: Microsoft Digital Crimes Unit and Resecurity on Storm-2990 and Storm-3119 (YAM Store), the Chinese-language phishing-kit and tenant-takeover ecosystem (S3), victims named in Japan, the United States, Australia and India, a takedown narrative with generic mitigations; the one concrete hunt (the PnP Management Shell service principal self-provisioned after a hijacked Global Administrator sign-in) concerns the broker's own hijacked tenants, not a decision for this constituency this week.
• borderline-drop: ARTEX, an open-source agentic penetration-testing framework abused against South Korean financial firms (S3; Zscaler, CrowdStrike), no constituency nexus, generic SQL injection and authentication bypass on peripheral web applications, and the store already carries several AI-agent intrusion entries.
• borderline-drop: SlugFest, npm Trusted Publishing slug takeover (S3; Bloom Security), fixed by GitHub on 2026-10-02, no exploitation, exposure only through suppliers that publish npm packages.
• borderline-drop: XWorm Reborn (Netskope, single source), CyberXero initial access broker (SOCRadar, nine days old, Ukrainian and Polish targets, the WordPress flaw already covered), UAT-6275 and TAG-193 (Talos with Recorded Future Insikt; CIS governments only, a generic WinRAR CVE-2025-8088 Startup-folder chain), Check Point on LiteLLM CVE-2026-35029 and CVE-2026-59823 (fixed 2026-04-17, authenticated, five LiteLLM entries already in the store) (S3).
• borderline-drop: Offerz, a Swiss voucher webshop whose January order-data copy in a less-monitored test environment was accessed on 2026-10-07 (S2), no public-sector nexus, no vector, incident floor; the transferable lesson (a supplier holding real data of a customer that has not yet moved onto its system) rides the PK Softech update through the PKZH notice.
• borderline-drop: Wolsfeld fire-service operations centre (Rhineland-Palatinate; single SWR relay of the prosecutors) and Mandurah SES (Western Australia; The Gentlemen listing, one server of pre-2022 records) (S4), no vector, no actor behaviour beyond impact, single source each, incident floor. SRF's blanket line that the Stadt Zürich fund is potentially affected (S2) is narrower in PKZH's own notice; the update carries PKZH's own statement and does not cite SRF.
• borderline-drop: NCSC.ch hub 13059, Citrix XenServer CVE-2026-79610, CVE-2026-102700, CVE-2026-102701, CVE-2026-79611 (S2, S1); host compromise needs a privileged guest, exploitation unknown. SAP October Patch Day (16 notes, none exploited; the one unauthenticated 9+ item is a Cloud SDK for Python library flaw), NGINX CVE-2026-97205, CVE-2026-93479, CVE-2026-15176 (configuration-dependent, high complexity), Gitea CVE-2026-59774 (fixed 2026-07-27, CVE published only now), AnyDesk "AnyPwn" (aged out), Siemens (third-party components and authenticated items), Ivanti Neurons for Patch Management and Xtraction (9.9 but authenticated), Oracle (the October CPU is dated 2026-10-20) (S1).
• borderline-drop: n0n maritime extortion batch, Ledger hardware-wallet implants, Femern Link Contractors (Fehmarnbelt tunnel contractor, no vector or actor; re-check next fire), Blueberry Pediatrics (a downstream Metabase victim, already covered), Egide, Medley At-Link, SpeeCAN, Park Dental, Joliet District 86 and the fuitesinfos claims (S4), out of nexus, claim-only or no behaviour.

Single-source: 2026-10-15/fake-watchguard-vpn-client-site-credential-capture is single-source (Huntress is the only source for this iteration; B2). Reduced-confidence inclusions: none.

Contradictions: the Danfoss improvement records that BleepingComputer reports no tie to the 2025 AK-SM 8xxA flaws while Censys points to them without Danfoss confirming the vector and NCSC.ch lists them as alleged; no source confirms a tie, and the entry says so. For the next audit: the verifier read Claroty's own page as saying the code-of-the-day gives a fully authenticated administrative session, as Censys does, so the older entry 2026-08-28/claroty-danfoss-ak-sm-800a-code-of-the-day-rce, which describes CVE-2025-41450 as an information disclosure, probably needs a correction record; this fire did not touch it.

Priority and ground calls: the Cisco gateway flaw is critical on the bar (exploited, unauthenticated root on the mail path, no workaround, compromised Secure Email Cloud tenants contacted by Cisco) although the fix ships as an automatic content update; PAN-OS is high (exploited, root, no workaround, but it needs credentials and management-interface reach); the Collabora flaw was first written notable for its single deployment mode and raised to high after verifier iteration 2: it is exploited, unauthenticated and critical, the web shells survive the update, and Collabora's own advisory says disabling the app does not stop it, so the decision this week is to update or remove it; the high bar asks for plausible exposure, and the store already carries Nextcloud entries for the public sector.

Candidate sources: cyberdaily-au added (S4). Not added: juniper-sirt and f5-security-advisories (S1; both vendor portals are JavaScript shells and no recipe exists, so they would only add standing repair orders), and Netskope Threat Labs and Bloom Security (S3; recipes unverified); the next audit may revisit them. blick-ch promoted to active per promotion_due although S2 judged it general news of limited depth; content_scope: general-news and a 48 h recipe note record that.

Rotation: the digest built, so no source was excluded as a recent attempt; slices ranked on the rotation cursor. S4's pool is 12 records and all were attempted by the previous two fires, so its ten oldest by cursor were allocated anyway.

Coverage gaps: inside-it-ch (article bodies behind the 429 checkpoint; RSS teasers only); fuitesinfos (listing shows 4 of 15 entries); swisspost-cybersecurity (marketing content); srf-news (feed and sitemap reach 36 to 48 h, so the 122 h lookback cannot be served); news-admin-ch (ignores date filters); kaspersky-securelist (no post since 2026-09-24); fortinet-fortiguard-blog (newest 2026-10-05); redcanary (blog moved to zscaler.com, own feed frozen); juniper-sirt (JSA pages are a JavaScript shell); reliaquest (sitemap lastmod bulk re-stamped, confirm every item by page date).

Operational notes: stop-hook nudges to commit arrived mid-run and were not acted on (nothing is committed before Phase 6). The harness offers no TodoWrite tool, so the phase plan lived in these notes. tools/source_health.py ran in Phase 5 and printed one UNSOLVED row, cnil-fr (url repointed, see sources_changed; CNIL's whole site answered a 503 maintenance page). One reader (jina) fetch was spent on a WatchGuard Secplicity page and one on the ENISA EUVD page inside extract's own fallback; neither is cited.

Verification: four independent cold passes (every ledger claim answered each time, 99 to 100 claims). Iterations 1 to 3 returned 12, 5 and 4 findings (the five of iteration 2 all came from the vendor's own Collabora advisory, which no earlier pass had read); iteration 4 returned one F3, a chained CISA citation in the PAN-OS entry. With one truth finding and no F1 or F4 the early exit applies: the fix was applied by the main agent against the saved Palo Alto page and the run publishes with verification_residual_count: 1, the final iteration's count. No CLEAN verdict was reached, so there was no confirmation pass; the Collabora entry (rewritten in iteration 2 and edited again in iteration 3) is the one the next audit should give an independent look. Declined findings: none after iteration 2 (the iteration 1 priority rebuttal was withdrawn in iteration 2).