Cases / UAI-2026-0091

Agent rewrote rule and skill files without approval, injected an ad-hoc macro that corrupted a UserForm, then falsely reported fixes.

Agent
Gemini CLI
Pattern
Ignored an explicit instruction
What was harmed
Local files and uncommitted work
Severity
Moderate: real damage that the user could repair, or money spent by mistake
Recovered
The report does not say
Reported
18 Oct 2026

Read the original report (a post on Qiita)

How sure are we?

Further right means the case carries more weight.

Evidence
B: A detailed self-report with commands, logs or steps that someone could check.
Who wrote the report
Written by the user
Was anyone watching?
Explicit limit, ignored: rule requiring approval before implementation
Vendor response
None found
Other trackers
Not in Manifold, Eon, AIR or METR

The guard that would have stopped it

Not written yet. See the guards we have.

Other cases like this

  1. 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
  2. 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
  3. UAI-2026-0081 Agent kept burning usage after cost-control instruction, exhausting weekly allowance with work unfinished and false completion claims. CodexIgnored an explicit instructionModerate, could have been severeEvidence B13 Nov 2026
  4. UAI-2026-0082 Voice agent sent email containing content the user explicitly excluded. CodexIgnored an explicit instructionModerate, could have been severeEvidence B13 Nov 2026
  5. UAI-2026-0077 Worked on wrong branch against repeated instructions: filed 42 issues, merged PRs, stashed user files. Claude CodeIgnored an explicit instructionModerate, could have been severeEvidence B12 Nov 2026