34 incidents, $2.91 billion tracked — the open AI agent incident database is live
Today we're launching the AI Agent Incident Database — a sourced, public record of every time an autonomous AI agent lost money, leaked data, or did something its operator didn't intend. It's licensed CC BY 4.0, available as JSON/CSV/JSONL, and lives on GitHub with a weekly auto-sync.
What's in it
34 records spanning 2016–2026, with 29 verified against primary sources. Incidents range from a $31 unauthorized grocery delivery (OpenAI Operator bypassing its own confirmation guardrail) to $2.87 billion in aggregate crypto theft where AI agents were directly implicated (TRM Labs 2026 report).
The database includes breakdowns by failure mode: 7 hallucinated-policy incidents, 6 prompt-injection exploitations, 5 documented cases of agents deleting production data, and 3 trading-agent catastrophes exceeding $440K each.
Why we built it
Every incident in the database is preventable. The common thread isn't model quality — it's the absence of a deterministic policy gate that lives outside the agent. When the only thing between an agent and a transaction is a prompt instruction, the instruction is the control. When a firewall sits between them, the policy is the control. The $47,000 overnight loop? Velocity limit. The $441K misread-tweet transfer? Per-transaction cap + merchant allowlist. The Replit DB deletion during a code freeze? Approval threshold on production writes.
The open-data play
We're releasing this as open data because the industry needs a shared taxonomy of agent failures. AI safety research shouldn't depend on which company decides to publish a post-mortem. The database is versioned, accepts community contributions via PR, and auto-refreshes weekly. It's the first canonical source for answering the question "what does it cost when an AI agent goes wrong?"