Skip to content

About

A structured incident database + post-mortem standard for agent failures.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

1 watching

Forks

agent-postmortems

A structured, public database of real AI-agent failures — and the schema for reporting them.

validate incidents release site code: Apache-2.0 data: CC-BY-4.0

Agents now take real actions in production — they call tools, spend money, write to databases, browse the web, and run code. They fail in novel but repeating ways, yet there is no shared, structured record of those failures. This repo is that record: one rigorous schema, a growing corpus of sourced incidents that conform to it, and the tooling to keep it honest.

→ Browse and search the corpus at swarmproof.github.io/agent-postmortems

What's in the corpus

| Incidents | 180 and growing (2023–2026) — see the live count badge above | | Type | 101 realized incidents · 79 demonstrated hazards / near-misses | | Coverage | 18 of 19 failure classes; 40 CVE-backed |

| Every record | conforms to one schema, cites public sources, classified against a versioned taxonomy |

Failures span prompt injection, tool misuse, data exfiltration, sandbox escapes, model-template poisoning, reward hacking, autonomous misuse, cost blowups, and destructive actions — across coding agents, MCP servers, browser agents, chatbots, and frontier-model evaluations.

How this differs

Several databases already catalogue AI harms: the AI Incident Database covers AI systems of every kind, AIAAIC collects AI, algorithmic and automation incidents, and the OECD AI Incidents Monitor documents AI incidents and hazards as an evidence base for policymakers. This corpus is narrower: it records only failures in which an agent took or recommended an action — a tool call, a write, a spend, a code execution, a commitment made to a user — not model outputs or predictions on their own. Each record carries an ordered causal chain of failure classes rather than a single label, so the sequence from trigger to outcome is part of the data. Every record is also published as machine-readable JSON validated against a published schema, and demonstrated hazards are admitted alongside realized incidents but labelled as such.

What a record looks like

Each incident is one YAML file (incidents/<incident_id>.yaml). Abbreviated example:

incident_id: 2025-replit-prod-db-deletion
title: "Replit AI agent deleted a production database during a declared code freeze"
date: "2025-07-18"
incident_type: incident          # incident (realized) | hazard (PoC / near-miss)
severity: high
confidence: confirmed
system:
  framework: Replit AI agent
  tools: ["database", "code-execution"]
  autonomy_level: supervised-autonomous
primary_failure_class: unsafe-action
failure_classes:                 # the causal chain, not a single label
  - {class: unsafe-action, subclass: data-deletion}
  - {class: excessive-agency, subclass: missing-approval-gate}
  - {class: hallucination, subclass: false-state}
trigger: "During an active code freeze, the agent ran destructive DB commands…"
root_cause: "Irreversible high-privilege actions against production with no enforced approval gate…"
blast_radius:
  data: {records: 1200, classification: confidential, description: "Production DB deleted…"}
  reversibility: partially-reversible
prevention: "Enforce dev/prod separation; require approval for destructive actions…"
mappings:                        # cross-references to external security frameworks
  owasp_llm: [LLM06, LLM09]      # Excessive Agency · Misinformation
  owasp_agentic: [T3, T5]        # Privilege Compromise · Cascading Hallucination
sources:
  - {url: "https://…", type: news, publisher: Fortune}

Required core + a large optional body (causation, attack_vector, timeline, cve/cwe, related_incidents, …). Records are cross-referenced to OWASP Top 10 for LLM Applications, OWASP Agentic AI Threats, and MITRE ATLAS wherever the crosswalk in schema/framework-mappings.yaml defines an equivalent for their failure classes; classes with no counterpart in those frameworks (such as autonomous-misuse and task-failure) carry no mappings block. Mappings are derived from the crosswalk, never hand-written. Full field reference: SCHEMA.md · taxonomy: TAXONOMY.md.

Standard stability: schema_version and taxonomy_version are both 1.0; additive optional fields and new classes do not bump the major version, while renaming or removing a field, class, or subclass does — and any taxonomy change bumps taxonomy_version and requires an ADR, per the governance note in TAXONOMY.md.

The one rule

Every incident needs at least one public, linkable source. No rumors, no speculation, no editorializing. Report failures factually; name systems without attacking them. Realized incidents and demonstrated hazards are both recorded, labeled honestly via incident_type. Disputed records are annotated, never deleted; a named party who contests a record can follow the process in DISPUTES.md. CI enforces the rigor — schema, taxonomy, id-uniqueness, link-liveness, and a neutrality lint all run on every PR.

Use the data

The corpus is regenerated into machine-readable JSON on every merge — no scraping, no API key:

BASE=https://raw.githubusercontent.com/swarmproof/agent-postmortems/main/export

# Whole corpus as one JSON array
curl -s $BASE/incidents.json | jq '.count'

# All prompt-injection incidents
curl -s $BASE/incidents.json | jq -r '.incidents[] | select(.primary_failure_class=="prompt-injection").incident_id'

# Everything CVE-backed
curl -s $BASE/incidents.json | jq -r '.incidents[] | select(.cve).incident_id'

# Incidents mapped to OWASP LLM06 (Excessive Agency)
curl -s $BASE/incidents.json | jq -r '.incidents[] | select((.mappings.owasp_llm // []) | index("LLM06")).incident_id'

The same from Python, standard library only:

import json, urllib.request
URL = "https://raw.githubusercontent.com/swarmproof/agent-postmortems/main/export/incidents.json"
data = json.load(urllib.request.urlopen(URL))
print(data["count"], [r["incident_id"] for r in data["incidents"] if r["primary_failure_class"] == "prompt-injection"][:3])

Releases: versioned snapshots of the three feeds (incidents.json, scenarios.json, seeds.json) are attached to tagged GitHub releases, so a downstream consumer can pin a tag instead of tracking main.

Two more feeds project the corpus for downstream tooling — scenarios.json (replayable chaos scenarios for stampede) and seeds.json (denial-of-wallet seeds for costbomb). Export entries carry natural-language trigger shapes, never runnable exploit payloads.

The site also offers a per-incident permalink with a copyable BibTeX citation and an RSS feed.

Contribute an incident

  1. Copy incidents/_TEMPLATE.yaml to incidents/<incident_id>.yaml (filename = incident_id, a YYYY-kebab-slug).
  2. Fill it in; classify with TAXONOMY.md.
  3. Validate: uv run scripts/validate.py, then uv run scripts/check_links.py incidents/<incident_id>.yaml — the scoped link check CI runs on a pull request. With no arguments, check_links.py checks the whole corpus.
  4. Open a PR — the template carries the sourcing + neutrality checklist; CI runs every gate.

New here? The good-first-issue label and the auto-filed candidate / coverage-gap issues are good starting points. See CONTRIBUTING.md.

Automated curation pipeline

Weekly workflows find new and not-yet-documented agent incidents and file them for review — automating the toil (discovery, de-dup, drafting, validation) but never merging anything without a human:

  • Discovery harvests candidates from security-research feeds, first-party lab/vendor disclosures, arXiv, Google News, and the NVD CVE / GitHub advisory databases, de-dupes against the corpus, and posts a ranked candidates issue.
  • Completeness critic diffs the corpus against curated external agent-incident lists and files a coverage-gaps issue — the "what did we miss?" review, automated.
  • Auto-draft (opt-in) has Claude research candidates and open review-ready PRs.

See docs/CURATION-PIPELINE.md.

Part of the Swarm Proof toolkit

Project What it does
agent-postmortems This repo — structured incident database + reporting standard
stampede Drives simulated agent traffic at a system under test
mockworld Mock external services (payments, email, exchange) for agent tests
mcp-probe Lint, contract-test, benchmark, and load-test MCP servers
costbomb Fuzzing for denial-of-wallet / unbounded-spend inputs
exactly-once Idempotency middleware for agent side-effects
cairn-protocol Checkpoint and recovery protocol for AI agents (ERC proposal, live on Base Sepolia testnet); the taxonomy's cairn_map aligns each class to its failure taxonomy
awesome-agent-reliability Curated list of agent-reliability resources

License & citation

Dual-licensed to separate tooling from data: code (schema, validators, scripts, site) under Apache-2.0; incident data (incidents/, export/) under CC-BY-4.0 — reuse freely with attribution. Cite via CITATION.cff, or use the per-incident BibTeX box on the site.

About

A structured incident database + post-mortem standard for agent failures.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages