What StupidLLM is

When an AI coding agent deletes a directory, leaks a secret, or rewrites half a codebase it was never asked to touch, that failure usually lives in a single tweet or bug report and then disappears. StupidLLM assigns those failures a stable identifier, scores them against a published rubric, and makes them searchable.

The database currently holds 124 documented incidents across 25 agents, plus 14 further multi-agent or unattributed reports — 138 in total. Every entry records the instruction given, the expected behavior, what actually happened, and the resulting damage — plus a source link so the claim can be checked.

What this is not

It is not a benchmark. Incident counts reflect public scrutiny as much as reliability, and there is no denominator of total tasks attempted. The methodology page states the limitations in full. Read the rankings as documented evidence, not as a defect rate.

License

The incident data is published under CC BY 4.0. You may reuse it, including commercially, with attribution.

Following new records

New records appear in the RSS feed, the full dataset is at /incidents.json, and the guardrail rules at /guardrails.json. To post new records into a Slack channel, add the feed URL with Slack’s RSS app. Records that are benchmarks or surveys rather than single failures are listed separately on studies.

How to cite

Each incident page also has a “Cite this record” panel with plain, Markdown and BibTeX versions.

Who maintains it

The StupidLLM team edits and maintains the database. We use AI research agents to find candidate incidents in public sources (GitHub issues, security advisories, vendor reports, news), and those agents are held to rules we audit: every record must link a public source, and no incident is published without one.

A record is labelled Source checked when the linked source was read and supports what the record says. Severity is the sum of four published factors (scope, recoverability, whether it actually happened, and reach beyond the user), shown on every incident page; all 138 were re-scored on that basis on 29 October 2026. Records we maintain in depth also carry who is affected, fix status, a sourced mitigation and a last-reviewed date.

Who pays for it

We take no money, sponsorship, affiliate commission or paid placement from the maker of any agent in the database. Sponsorship is open only to security, backup and recovery tools, is labelled as such, and gives the sponsor no say over records, scores, rankings or coverage. The same rules apply to our paid services: a review or a data license buys work and terms, never a change to a record.

The Agent Guardrail Pack linked from some incident pages is sold by agentkitworks, which is run by the same people as StupidLLM. Every incident page that links to it also shows free rules for the same commands.

Corrections

If an entry is wrong, email hello@stupidllm.com with the incident ID, or use the correction link on any incident we maintain in depth. We correct the record or withdraw it; we do not leave a disputed claim standing unchecked. A vendor that disputes a record gets a reply before anything else in the queue, and if the record stands, a link to the vendor’s published response is added to it. To add an incident, report it here.