What has already happened

Responsible AI Collaborative

Real-world AI incidents (harms already caused) across all AI system types, from public reporting and community submission.

Scale:

~1,600 unique incidents (IDs through #1618 in the May to July 2026 roundup, published 11 August 2026)

Agentic:

General AI, not agent-specific, but increasingly includes agentic-system incidents as a subset within its taxonomy.

Use for:

A single searchable, community-reviewed report on a specific documented harm.

OECD.AI Policy Observatory

AI incidents and hazards (near-misses / plausible-harm events) detected from global news media via an automated news-intelligence pipeline.

Scale:

16,000+ incidents and hazards combined as of 11 July 2026 (per OECD's taxonomy, 9,000+ incidents / 5,000+ hazards)

Agentic:

General AI; media-detection means agentic incidents are captured only when reported in mainstream/trade press, not systematically tagged as agentic.

Use for:

Scale and trend analysis; the largest, most automated feed.

Charlie Pownall (independent, volunteer-run public-interest project)

AI, algorithmic and automation incidents and controversies, including reputational, ethical and governance failures beyond strict harm events.

Scale:

Entries running past #2264 as of mid-2026 (no official running total published)

Agentic:

General AI/algorithmic; its broad scope picks up agentic-tool controversies that harm-focused trackers may exclude.

Use for:

Controversy and governance context around an incident, not just the technical failure.

MIT AI Risk Initiative

Reclassifies AIID's raw reports against MIT's own risk taxonomy and a harm-severity scale, with EU AI Act risk-level tagging.

Scale:

Classifies 1,400+ incidents sourced from AIID; June 2026 update focused on classifier validation

Agentic:

General AI; taxonomy-driven, though its causal/domain tags allow filtering toward agentic-system failures.

Use for:

An incident pre-classified by severity, domain and EU AI Act risk level.

Damien Charlotin (independent legal researcher, HEC Paris)

Court and tribunal decisions worldwide where a party was found to have relied on AI-fabricated content (invented citations, false quotes).

Scale:

Large and growing weekly; the site is the authority for the current count, so we do not hardcode a figure

Agentic:

Litigation-specific, for one failure mode (fabricated legal citations), increasingly involving AI legal-drafting tools.

Use for:

The go-to citation for legal/litigation risk from generative AI.

METR (AI evaluation nonprofit)

Documented incidents in which AI agents deliberately acted against their users' intentions, scored on two axes: overreach and deception.

Scale:

Interactive chart; last updated 19 May 2026

Agentic:

Agent-specific by construction; the only tracker here built around agent behaviour rather than AI harm in general.

Use for:

Cases where an agent went beyond, or hid, what it was asked to do.