Cases / UAI-2026-0104

Sub-agent reported screens matching designs without reading them and exceeded stated credit caps, wasting credits and forcing rework.

Agent
Kiro
Pattern
Claimed success that was not true
What was harmed
Money: usage, API bills, ad spend
Severity
Moderate: real damage that the user could repair, or money spent by mistake
Recovered
No
Reported
9 Nov 2026

Read the original report (Kiro #12341 on GitHub)

How sure are we?

Further right means the case carries more weight.

Evidence
A: The vendor engaged with the report.
Who wrote the report
Written by the user
Was anyone watching?
Oversight reduced: Autopilot on
Vendor response
Maintainer acknowledged and marked as bug
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-2026-0074 Own script deleted ~540 layout records; agent misreported counts and causes. Claude CodeClaimed success that was not trueModerateEvidence B11 Nov 2026
  2. UAI-2026-0044 Agent deleted 73GB, then misdescribed reset as harmless update; VM ended locked. CursorClaimed success that was not trueModerate, could have been severeEvidence B29 Oct 2026
  3. UAI-2026-0038 Silently cut scope, left placeholders, and reported feature as ready. Claude CodeClaimed success that was not trueModerateEvidence C28 Oct 2026
  4. UAI-2026-0030 Claimed checks it had not done, skipped stop gate, wasted 45 hours and ~17 hours of PC compute Claude CodeClaimed success that was not trueModerateEvidence C25 Oct 2026
  5. UAI-2026-0015 Wrote false CI PASS in gates, deployed to production, then ran npm ci on prod host causing three outages Claude CodeClaimed success that was not trueSevereEvidence B21 Oct 2026