Patterns

Did something nobody asked for

The agent took an action outside the task: a deploy, a test run against production, a purchase.

19 cases, 2 of them severe.

Guards

Cases

  1. 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
  2. UAI-2026-0075 Agent dropped Supabase schema still exposed to PostgREST, causing production API outage. Claude CodeDid something nobody asked forModerateEvidence B12 Nov 2026
  3. UAI-2026-0078 Agent read another project's .env and echoed full API token in chat. Claude CodeDid something nobody asked forModerateEvidence C12 Nov 2026
  4. UAI-2026-0072 Repeated paid test runs against production API drained balance, taking user's app AI features offline. Claude CodeDid something nobody asked forModerateEvidence B10 Nov 2026
  5. UAI-2026-0071 Browser automation failed over to Chrome on a different machine without asking. Claude CodeDid something nobody asked forModerateEvidence B9 Nov 2026
  6. UAI-2026-0070 Agent sent an email to a vendor without user approval; later apology sent. Claude CodeDid something nobody asked forModerateEvidence C8 Nov 2026
  7. UAI-2026-0066 Agent moved 11 GB of files to Recycle Bin without any visible approval; bin then emptied. Claude CodeDid something nobody asked forModerateEvidence B6 Nov 2026
  8. UAI-2026-0053 Asked only to reboot a server, agent edited fstab, broke systemd and containers, ran random scripts. CodexDid something nobody asked forSevereEvidence A3 Nov 2026
  9. UAI-2026-0049 Published paid Meta ad campaign with budget typo (300150) unconfirmed; spent about USD 148. Claude CodeDid something nobody asked forModerateEvidence B1 Nov 2026
  10. UAI-2026-0047 COM script quit user's already-open PowerPoint, losing an hour of unsaved work. Claude CodeDid something nobody asked forModerateEvidence B31 Oct 2026
  11. UAI-2026-0042 Installed one project's hooks at user scope and parent folder, affecting all sessions for 15 days. Claude CodeDid something nobody asked forModerateEvidence B29 Oct 2026
  12. UAI-2026-0039 Misread 'let me do it' and archived user's public GitHub repo. Claude CodeDid something nobody asked forModerateEvidence B28 Oct 2026
  13. UAI-2026-0040 Codex silently edited stale same-named files from another directory and reported completion. CodexDid something nobody asked forModerateEvidence B28 Oct 2026
  14. UAI-2026-0035 Agent used TTS key for paid Vertex image/video calls, about IDR 1.07M billed, continued after cost question. Claude CodeDid something nobody asked forModerateEvidence B27 Oct 2026
  15. UAI-2026-0027 Agent used keychain credentials to log in to third-party BBS and made unasked connections. Claude CodeDid something nobody asked forModerateEvidence B24 Oct 2026
  16. UAI-2026-0019 Unrequested git checkout and reset --hard destroyed gitignored .env with API keys. Claude CodeDid something nobody asked forSevereEvidence B22 Oct 2026
  17. UAI-2026-0016 Agent made unrelated change and deployed to api twice without approval. Claude CodeDid something nobody asked forModerateEvidence C21 Oct 2026
  18. UAI-2026-0012 Merged PR in unrelated repo, re-added disabled Trivy scan, and misreported rule compliance. Claude CodeDid something nobody asked forModerateEvidence B20 Oct 2026
  19. UAI-2026-0006 Changed process step so records were posted unapproved; also edited user's rules file unasked. Claude CodeDid something nobody asked forModerateEvidence B18 Oct 2026