Skip to content

agent-postmortems corpus — 2026.11.1 (148 incidents)

Choose a tag to compare

@MarouaBoud MarouaBoud released this 17 Nov 14:52

A versioned snapshot of the agent-postmortems corpus — a structured, public database of real AI-agent failures.

This snapshot

  • 148 incidents (91 realized · 57 demonstrated hazards / near-misses), 2023–2026
  • 17 of 19 failure classes represented; 27 CVE-backed
  • Every record conforms to the v1 schema and is classified against the versioned taxonomy

Since 2026.11

  • 67 new records (39 realized · 28 hazards) covering the July–November 2026 backlog: seven operations from Anthropic's September threat report and the GTIG, Unit 42, Storm-3168, Thai Finance Ministry, PaperCut and Hermes agent-driven campaigns; the OpenAI, AISI, Google and DeepSeek evaluation and research disclosures; the Claude Code, Codex and Gemini destructive coding-agent incidents; the LiteLLM, Cline, MemOS, Apex Accelerator and skills.sh supply-chain compromises; and 28 demonstrated hazards against coding assistants, agentic browsers and agent frameworks (Kiro, Copilot, Azure DevOps MCP, Rovo, Manus, Flowise, Langflow, MCPHub, Paperclip, GitSpawn)
  • Anthropic eval-escapes record corrected against the September primary source, adding the fourth case
  • Link-liveness gate now checks only the records a PR changes and treats an unreachable source with a live archive as a warning, so a flaky third-party host no longer fails unrelated PRs

Assets

  • incidents.json — the full corpus as one JSON array
  • scenarios.json — replayable scenario shapes (for stampede)
  • seeds.json — denial-of-wallet seeds (for costbomb)

Export entries carry natural-language trigger shapes, never runnable exploit payloads.

Use the data

Incident data is licensed CC-BY-4.0 — reuse freely with attribution. Cite via CITATION.cff.