[Opus 4.6 in Claude Code] Misleading / lying behaviour
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Claude modified files I didn't ask it to modify
What You Asked Claude to Do
I asked Claude to run an isolated test of a multi-step data processing pipeline to verify that a bug fix was working before running the full system. The pipeline has multiple stages, each taking 30-120 seconds to execute.
What Claude Actually Did
Claude presented cached test results as fresh validation without disclosing that the test didn't actually execute. The 0.0s execution time was a clear signal that the code path wasn't exercised, but Claude reported "3 divisions found" as if the test confirmed the fix worked. This led to a live run that failed, wasting time and compute credits
This is a reasoning/honesty failure, not a tool limitation — the information to detect the problem (0.0s runtime) was right there in the output.
Expected Behavior
- When a test step completes in 0.0 seconds while identical steps take 60-120 seconds, Claude should flag this as anomalous and investigate rather than silently reporting the result as valid.
- When the user explicitly asks to "test whether the fix works," Claude should ensure the test actually exercises the fixed code path — not load a pre-fix cached result and report it as post-fix validation.
- If Claude detects that a result may come from caching, it should disclose this: "This result was loaded from cache (0.0s) — it may not reflect the current code. Clear the cache and re-run to verify the fix."
- Claude should never present unverified results as verified. This is the same class of error as stating a hypothesis as fact — the information to detect the problem was in Claude's own output, and Claude chose to report confidence rather than uncertainty.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Yes, every time with the same prompt
Steps to Reproduce
_No response_
Claude Model
Opus
Relevant Conversation
Impact
Critical - Data loss or corrupted project
Claude Code Version
2.1.104
Platform
Other
Additional Context
_No response_
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗