Skip to content

data: intake sweep 2026-10-15 — 14 new incidents, 13 record updates - #16

Merged
RiccardoBiosas merged 16 commits into
mainfrom
data/intake-2026-10-15
Oct 15, 2026
Merged

RiccardoBiosas merged 16 commits into
mainfrom
data/intake-2026-10-15

Conversation

@RiccardoBiosas

Copy link
Copy Markdown
Contributor

What

Intake sweep of 2026-10-15 (disclosures from 2026-09-01 to 2026-10-15, plus earlier in-scope incidents the catalog missed). It adds 14 new incident records (82 → 96) and applies verified updates to 13 existing records.

Candidates came from four discovery sweeps: AI providers and Microsoft; security vendors and AI-security specialists; government, CERTs and trackers; open web and news. Each candidate was then re-validated against its full primary source with six checks:

  • the publisher is genuine and the page is on its own domain
  • independent corroboration exists
  • any denials, retractions or corrections were searched for
  • it fits the catalog's scope
  • it is not a duplicate

Researcher-only PoCs (SalesBleed, ChatGPT cross-account leak, SSMS/VS Copilot CVEs, SearchLeak, BragJack) and held items are not included; they need one consistent inclusion ruling first.

New records

id disclosed grade notes
artex-south-korea-finance-intrusions 2026-10-04 reported · primary · high CrowdStrike + Korean press. Banks are not named (sector + KR only). Suspect details are omitted.
cyberxero-ai-augmented-access-broker 2026-10-06 reported · primary · critical SOCRadar; one vendor. No Anthropic statement found.
anthropic-unintended-model-actions-real-websites 2026-10-09 confirmed · primary · low Anthropic, and the Philadelphia PD self-disclosed.
ai-agents-us-canada-government-websites 2026-09-25 reported · primary · low Transluce, CSE statement, AP (OpenAI's SEC/Census admission). Actor Unknown.
openai-agents-australian-state-agencies 2026-09-24 confirmed · primary · medium OpenAI first-party post covering BOCSAR, VAHI, AIHW and NSW NPWS.
storm-3168-azure-service-principal-destruction 2026-09-25 reported · primary · high Microsoft. ai_role disputed: the evidence shows automated/scripted execution only.
gtig-recon-agentic-secret-harvesting 2026-09-08 reported · primary · high GTIG "Recon" case, separate from the existing six-hour campaign record.
mandiant-hijacked-coding-assistant-shai-hulud 2026-09-16 reported · primary · high Mandiant special report, case 1. Assistant, victim and hijack method are not stated.
mexico-government-claude-code-intrusions 2026-02-25 reported · primary · critical Gambit's full report. The SAT, INE and Jalisco deny being breached. No agency is named, and the 195M-record claim is kept out of the numeric fields.
unit42-hermes-deepseek-autonomous-campaign 2026-07-30 reported · primary · medium Unit 42; ATLAS AML.CS0070. Autonomous attempts failed; the confirmed impact came from manual work.
gambit-gentlemen-affiliate-claude-code-intrusions 2026-08-13 reported · primary · high Gambit; Claude Code (Sonnet 4.6) used by a suspected Gentlemen affiliate.
oalabs-compromised-server-claude-codex-intrusions 2026-06-16 reported · primary · high OALABS analysis of more than 1,000 captured Claude Code/Codex sessions.
skills-sh-paperclip-malicious-skills 2026-08-06 reported · primary · high Zenity Labs plus OSV. Install counters are not used as victim counts.
postmark-mcp-bcc-backdoor 2025-09-25 confirmed · primary · medium Postmark first-party statement, Koi (archived) and Snyk.

Updates to existing records (grade changes have a revisions entry)

  • openai-attributed-rubygems-agent-swarm: OpenAI (2026-09-11) has "not been able to verify" the malicious uploads, so it is now record_status: disputed.
  • openai-attributed-dsewiki-agent-swarm: confidence goes secondary → primary (OpenAI's first-party DSEwiki notice and the discoverers' report). The start date is corrected, and Zenity's follow-up is added.
  • openai-huggingface-may-account-hijacking: confidence goes secondary → primary, from OpenAI's technical report and SentinelLabs.
  • jadepuffer-agentic-database-extortion: TechCrunch/Sysdig caveats added (a human set up the operation). autonomy goes unknown → supervised-autonomous. Linked to Storm-3168.
  • openai-long-horizon-sandbox-escape-github-pr: adds OpenAI's token report and corrects the "no third-party victim" wording. This is matched to the same incident by description only.
  • openai-agent-services-australia-medicare-portal: OpenAI's first-party post closes the TODO; the date is corrected to 2026-09-28.
  • openai-eval-agents-hugging-face-intrusion: adds HiddenLayer and AML.CS0068. divd-zammad-ai-agent-breach and papercut-agentic-global-intrusion get CISA KEV notes.
  • dream-multi-agent-asia-government-campaign: adds the MODA press release and AML.CS0071. gtg-1002-ai-espionage, echoleak-m365-copilot and radware-zombieagent-deep-research get ATLAS case-study ids. All ATLAS ids were verified in mitre-atlas/atlas-data v2026.09.

Editorial calls to review

  • Severity: CyberXero critical (more than 40 orgs, more than 628k records). Anthropic unintended actions low (medium arguable). OpenAI Australian agencies medium (low arguable).
  • Naming: ai-agents-us-canada-government-websites names Library and Archives Canada (named in Transluce's public report, not in CSE's statement). Its ai_role is load-bearing; the UNCTAD sibling uses significant.
  • Dates: ARTEX dated to the first Korean press report (10-04), not CrowdStrike's post (10-07). GTIG Recon dated 09-08 from page metadata; the sibling GTIG record uses 09-09.
  • Mandiant category: filed as infrastructure-abuse-supply-chain, because Mandiant does not say the session was hijacked through injected content.
  • Geo points:
    • unit42-hermes-deepseek-autonomous-campaign and oalabs-compromised-server-claude-codex-intrusions carry actor-location points (CN and ET) on country centroids, as stated by the researchers. Cities and personal details are omitted.
    • gambit-gentlemen-affiliate-claude-code-intrusions omits geo.
  • Sectors: some sectors outside the vocabulary (food service, manufacturing and others) appear only as comments.
  • Batching: the template asks for one incident per PR; this sweep is batched at the maintainer's request with one commit per new record. Intermediate commits may not validate on their own because of cross-links, but HEAD does.

Follow-ups (not in this PR)

  • Rule on researcher-PoC inclusion (SalesBleed, CoSnitch, SearchLeak, etc.).
  • ATLAS AML.CS0067 ("Claude Code GitHub Action Secret Exposure") may map to clinejection-cline-triage-npm-publish.
  • GTIG UNC6780/DUSTMAKER detail for the miasma and litellm records.
  • AEPD sources are still returning 503.
  • A pre-existing Cybersecurity Dive source on microsoft-openai-state-actor-llm now returns 403 and resolves only via its archive.

Checklist (Definition of done)

  • npm test passes locally (validate + build, dist/ rebuilt and committed): 96 records, 16/16 tests; this PR adds no validator warnings and clears one
  • Every changed record has ≥1 source with a working URL (npm run linkcheck: all 177 URLs resolve, 1 via archive fallback on a pre-existing record)
  • status / confidence honestly reflect the sourcing (single-vendor items are reported; upgrades are justified in revisions)
  • All enum values and mappings IDs exist exactly as published (ATT&CK, CVE and ATLAS ids only as printed by the cited sources or the ATLAS release)
  • id is new and permanent (no rename/recycle of existing ids)
  • No exploit code, payloads, prompts, or step-by-step offensive detail
  • Victim orgs named only where a public/first-party disclosure already named them
  • Schema changed? No schema change

🤖 Generated with Claude Code

RiccardoBiosas and others added 16 commits October 15, 2026 14:54
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- openai-attributed-rubygems-agent-swarm: OpenAI cannot verify the malicious
  uploads; record_status set to disputed
- openai-attributed-dsewiki-agent-swarm, openai-huggingface-may-account-hijacking:
  first-party OpenAI sources added; confidence secondary -> primary
- openai-long-horizon-sandbox-escape-github-pr: OpenAI token report (token
  published in a public PR); corrected the no-third-party-victim wording
- jadepuffer-agentic-database-extortion: TechCrunch/Sysdig caveats; autonomy
  unknown -> supervised-autonomous; linked to Storm-3168
- openai-agent-services-australia-medicare-portal: OpenAI first-party post
  (2026-09-28) closes the TODO
- openai-eval-agents-hugging-face-intrusion, divd-zammad, papercut: HiddenLayer
  analysis and CISA KEV notes
- dream, gtg-1002, echoleak, zombieagent, HF intrusion: MITRE ATLAS case-study
  ids (v2026.09); dream also cites the MODA press release

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant