Incident database

AI agents break things.
Each failure gets an ID.

Like CVE, for AI coding agents. Every case where an agent caused real damage gets a STUPID-ID, a severity score against a published rubric, and a primary source you can check — 82 so far across 23 agents, one added most days. All of it queryable from inside your agent.

Browse 82 incidentsGet incident alerts

How this is built

Verified against source
79/82
Added in 14 days
15
Most recent
2026-08-30
Licence
CC BY 4.0

Every incident is scored against a published rubric and carries a primary source. Where those sources come from: 31 news report, 19 github issue, 13 github pr, 9 benchmark, and 4 other kinds. Counts here measure what has been documented, not how often an agent fails — see the limitations.

Three ways to use it: browse, clone the corpus (every record’s history is a dated commit), or query it from inside your agent.

82
documented failures
6.6
avg severity /10
23
vendors tracked
8
ranked agents

Most documented agents

Counts reflect how much a tool is used and reported — not how often it fails. Claude Code has the most documented incidents, at an average severity of 7.3. Gemini CLI's incidents average 9.4.

Claude Code21avg 7.3Devin10avg 3.9Cline5avg 5.5Gemini CLI4avg 9.4GitHub Copilot4avg 8.2Aider3avg 4.5Claude3avg 3.6Cursor3avg 7.6

15 further agents are excluded — fewer than 3 documented incidents is not enough to rank on.

Latest reports

6.5
Claude Code destroyed a production Vultr server while the user was actively typing "don't destroy it" (GitHub
3.5
OpenAI Codex deleted important project files with no explicit request or confirmation, exact command never captured (GitHub #38312)
5.4
Claude Code loaded managed config from a world-writable Windows path, enabling local privilege escalation (CVE-2026-35603)
7.3
Claude Code overwrote live site pages with no backup during an "emergency" takedown, permanently losing 4 days of uncommitted work (GitHub #88384)
9.2
Claude Code sub-agent's recursive delete escaped its working directory and wiped an entire Windows user profile — 234,884 files — after the safety classifier failed open (GitHub #86872)
7.8
Claude Code hand-ran an unrequested lftp mirror --delete against a production host, wiping a customer photo-uploads directory, then reported a failed deploy as successful (GitHub #89014)
3.5
Claude Code silently deleted its own shared task-store files ~5 seconds after a teammate agent finished the last task, four times, with no Task tool call in the logs (GitHub #88346)
8.3
Self-propagating npm worm plants persistence in Claude Code's hook files to survive credential rotation (ChainDrop / keyv supply-chain attack)

Severity distribution

Distribution of 82 incidents by severity score (0–10)
  • Critical 9–10
  • High 7–8
  • Medium 4–6
  • Low 0–3

Most common failure modes

Distribution of failure modes across all documented incidents.

Highest severity on record

STUPID-2026-0027
10.0critical

Gemini CLI silently executed arbitrary code from an untrusted repo (CVE-2026-12537, CVSS 10.0)

What is StupidLLM?

According to StupidLLM's incident database, 69 AI agent failures have been documented across 23 agents, plus 13 further multi-agent or unattributed reports — 82 in total — with an average severity of 6.6/10. Every incident is severity-scored using a CVSS-inspired rubric, verified against source evidence, and searchable by agent, failure mode, and root cause.

How are AI agent incidents scored?

Every incident is scored on a 0–10 scale. Scores of 9–10 are critical, 7–8.9 high, 4–6.9 medium, and below 4 low. The full rubric is on the methodology page.

Which AI coding agent has the most failures?

Claude Code has the most documented incidents (21), at an average severity of 7.3/10 — while Gemini CLI has the highest average severity at 9.4/10, with 4 documented incidents of its own. Incident counts reflect public scrutiny and adoption as well as reliability. Agents are only ranked once they reach 3 documented incidents. See the rankings above.

MCP server

Query this from inside your agent

Every incident here is callable over MCP — search by agent, failure mode or severity, and cite the source directly.

claude mcp add --transport http stupidllm https://www.stupidllm.com/api/mcp/

Read-only · no key · no signup

What you get back, and how to cite it

Incident alerts

Get told when an agent breaks something

We document AI agent failures daily, severity-scored against a published scale. When one lands at 7.0 or above — deleted data, leaked secrets, broken production — you get an email with the source. When nothing does, you get nothing.

Confirmation required. Unsubscribe in one click, from any email.