AI agents doing what nobody asked them to.

We read the public bug reports where people describe AI agents wiping drives, running up bills and breaking production. We grade the evidence, group the cases by how they happen, and publish the guard that would have stopped each one.

How 3,229 search hits became 87 cases

17 October to 16 November 2026. Bars are drawn to scale. Almost everything we read gets thrown out.

  1. 3,229
    Search hits. Public reports that match phrases like "deleted", "without permission" or "data loss".
  2. 1,387
    Candidates. We drop feature requests and reports with no harmful action in them.
  3. 168
    An agent caused it. We read each report in full. An AI agent decided to act, and the user did not want the result. No attacker, no jailbreak. Plain product bugs are out.
  4. 87
    Something real broke. Lost files, an outage, money gone, a message sent. Minor slips are out. These are the cases on this site.

These 87 are a floor. Most people never file a report, and we only search Claude Code, Codex and OpenClaw issue trackers on GitHub, plus the Cursor forum. Read the method

How sure we are

Every case is a public report, and no vendor has confirmed any of them. So we record three things that tell you how much weight a case can carry.

Evidence grade

  • A: vendor engaged 3
  • B: detailed, checkable 63
  • C: thin, or the agent's word 21

Who wrote the report

  • The user 46
  • The user, quoting the agent 4
  • The agent itself 37

Was anyone watching?

  • Oversight on 6
  • Not stated 45
  • Oversight turned down 36

How it happens

Latest cases

  1. UAI-2026-0085 Overnight parallel agent builders and reviewers consumed about 67% of weekly Max allowance Claude CodeRan away in a loopModerateEvidence B16 Nov 2026
  2. UAI-2026-0086 Resumed agent kept expanding diagnostics despite timeboxes, burning 1.18B tokens without deliverables. CodexRan away in a loopModerateEvidence B16 Nov 2026
  3. UAI-2026-0087 Agent duplicated work, added unrequested files, used disabled browser; user reports ~75,000 credits and three quota resets. CodexDid something nobody asked forModerateEvidence B16 Nov 2026
  4. UAI-2026-0084 Exceeded stated rate limits on paid API and overwrote stored response; provider disabled account, no refund. Claude CodeIgnored an explicit instructionModerateEvidence B15 Nov 2026
  5. UAI-2026-0083 Subagent rsync --delete with empty variable wiped WSL root and mounted Windows drive folders. Claude CodeRan a destructive commandSevereEvidence B14 Nov 2026
  6. UAI-2026-0080 Agent spawned live Codex runs despite prohibition; also introduced restore step discarding user edits. Claude CodeIgnored an explicit instructionModerateEvidence C13 Nov 2026
  7. UAI-2026-0081 Agent kept burning usage after cost-control instruction, exhausting weekly allowance with work unfinished and false completion claims. CodexIgnored an explicit instructionModerateEvidence B13 Nov 2026
  8. UAI-2026-0082 Voice agent sent email containing content the user explicitly excluded. CodexIgnored an explicit instructionModerateEvidence B13 Nov 2026
  9. UAI-2026-0075 Agent dropped Supabase schema still exposed to PostgREST, causing production API outage. Claude CodeDid something nobody asked forModerateEvidence B12 Nov 2026
  10. UAI-2026-0076 Agent filled system drive with ~1TB temp builds, saturated CPU, took unauthorized actions. Claude CodeRan away in a loopSevereEvidence C12 Nov 2026

See all 87 cases