@pipeworx/ai-incident-db

Connect: https://pipeworx.io/mcp — every tool in the catalog, including @pipeworx/ai-incident-db’s. Install: one-click buttons

No MCP client? Skip the connection: POST https://gateway.pipeworx.io/v1/tools/search_packs {"query":"..."} to find a tool below, GET /v1/tools/<name> for its schema, POST the same URL with arguments for the data — see For AI agents.

Tools: 4

AI Incident Database (AIID), maintained by the Responsible AI Collaborative: a curated record of real-world harms caused by deployed AI systems, with the organisations that developed and deployed the system, the parties harmed, and links to the reports about each incident.

Stored copy. The publisher’s data API now requires a logged-in account (checked 2026-11-15), so this pack answers from a stored copy of the publisher’s weekly public database snapshot. Every answer carries data_as_of (the snapshot date), answered_from: "stored_copy", the source and the licence attribution. There is no live fallback: until the first load, every tool answers found: false, reason: "stored_copy_unavailable" and says so, which is not an empty result.

Tools

• search_incidents(query?, start_date?, end_date?, limit?, offset?): every word must appear in the incident title or description; if no incident has all words, incidents with any word are returned with a note. A search with no hits returns found: false, reason: "no_match".
• get_incident(incident_id): one incident with deployer, developer and harmed parties and its linked reports (title, source domain, link, publication date, language). A merged incident answers reason: "merged" with the number it was merged into.
• ai_incident_db_list_recent(limit?): newest incidents by incident date.
• list_taxonomies(): the classification schemes (CSETv1, CSETv0, GMF, MIT and annotator variants) with weight, field count and a short description.

Data source

AI Incident Database, Responsible AI Collaborative, weekly full-database snapshot. About 1,800 incidents, 7,600 entities and 7,000 linked reports in the 2026-11-09 snapshot.

Licence. The publisher’s Terms of Use, section “Data License”, licenses the incidents, quickadd, duplicates, taxa, classifications, entities and entity_relationships collections under Creative Commons Attribution-ShareAlike (the snapshot’s own notice says CC BY-SA 4.0). Reuse of this pack’s incident, party and taxonomy records must credit the AI Incident Database and carry the same licence. The report text and the submissions collection are not licensed that way and are not stored. Report titles, links and dates point to third-party publications that keep their own rights.

What is left out. Report authors and submitters, editors and every account field, editor notes, report text, images, similarity lists and embeddings are not stored or served, and e-mail-like strings in free text are replaced. Party names are the publisher’s entity names; incident descriptions are the publisher’s own text and may name people the incident concerns.

Refresh

scripts/ingest-ai-incident-db.mjs downloads the newest snapshot linked from the publisher’s public snapshots page, writes about 22 objects (about 5 MB) to the shared datasets bucket under ai-incident-db/, writes _meta.json last, and reads it back. --dry parses and prints counts without writing. Weekly via .github/workflows/ai-incident-db-refresh.yml.

Tools

• search_incidents — Search the AI Incident Database (stored copy of its weekly public snapshot, dated in every answer). Combines free-text with date range filters. Returns incident IDs, titles, dates, and descriptions, n
• get_incident — Fetch a full AI Incident Database record by numeric incident_id from the stored snapshot copy; returns title, date, description, alleged deployer/developer/harmed parties, and all linked report titles
• ai_incident_db_list_recent — Most recent incidents by incident date, from the stored snapshot copy.
• list_taxonomies — Taxonomies used to classify incidents (namespace, weight, field count, short description), from the stored snapshot copy.