How we track
Existing AI incident trackers draw on news stories, lab disclosures and security research. Most agent harm never reaches any of those. It shows up in bug reports, filed by the person it happened to. That is where we look.
What counts as a case
- An AI agent decided to act. It ran a command, sent a message, spent money or changed a system. A plain product bug does not count.
- The user did not want that outcome. No attacker, no jailbreak, no prompt injection.
- Something real broke. Lost files, an outage, money gone, a message sent. We list minor slips separately and leave them off the site for now.
Where we look
The public issue trackers for Claude Code, Codex, OpenClaw and Gemini CLI on GitHub, and the Cursor community forum. Gemini CLI produced no cases in our first window. We do not yet cover Reddit, Hacker News or X.
How we find and read cases
- We search each source with about 30 harm phrases such as "deleted", "without permission" and "data loss". The first 30-day window gave 3,229 hits.
- We drop feature requests and items with no harmful action in them. That left 1,387 candidates.
- Claude reads every candidate in full, including the first replies, and labels it. Each "real" label needs a quote from the report as evidence.
- A person settles every disagreement and reviews every case before it is published.
- We merge duplicates. One user who filed 130 near-identical complaints counts once.
The first window found 168 real cases. 87 caused material harm, and those are the cases on this site.
How we grade the evidence
- A The vendor engaged with the report. (3 cases)
- B A detailed self-report with commands, logs or steps that someone could check. (63 cases)
- C A thin report, or one that rests on the agent's own account. (21 cases)
Who wrote the report
Many reports are written by the agent itself, at the user's request. In 37 of 87 cases, the agent wrote or likely drafted the report. We mark a report as agent-drafted when it says so, or when it has the agent's voice: a first-person account of "my mistakes", a structured incident summary, or text the user says came from the session. An agent's account of its own failure is still evidence, but weaker evidence, and we say so on every case.
Oversight
We record whether the user had turned oversight down: auto mode, no approvals, or an unattended run. 42 reports state a mode. In 36 of those, oversight was turned down.
What we do not claim
- No vendor has confirmed any case. These are self-reports.
- The counts are a floor. Most people never file a report, and keyword search misses some that do.
- Counts by agent show where people report, not which agent is safer.
Privacy and corrections
We link to the public report, but we do not repeat the names of the people involved, or of their apps or companies. To correct a case or ask us to remove one, contact us.