[Bug] Anthropic API safety filter: False positive on defensive security code hardening, silently downgrades model without diagnostic feedback
Bug Description
A routine defensive-security coding session was flagged, downgrading Fable 5 to Opus 4.8.
I was hardening my own Claude Code plugin against a prompt-injection hole — a component that reads untrusted session-transcript data could be tricked into executing instructions embedded in it. Legitimate, defensive, on my own code. Over the session the conversation naturally contained security-testing vocabulary: describing the injection, an "override"-style adversarial test string, an unsupervised process writing files, and a test that verified the fix blocked the attack by confirming no side-effect file was created.
The safeguard fired and silently switched me to Opus 4.8 mid-conversation. Nothing in the work was unsafe — it was exactly the kind of "safe and routine cybersecurity work" the notice admits may be flagged. The flag also gave no indication of what triggered it, so I couldn't tell whether I'd done something wrong or hit a false positive; that ambiguity is its own cost.
Two pieces of feedback:
- This is a clear false positive on defensive security work. Building and testing an injection fix requires writing about attacks; a filter that reacts to that vocabulary without reading intent will keep catching the people doing the right thing.
- When a switch happens, some signal about the category that triggered it (even coarse) would help users judge whether to adjust course or just file feedback.
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.214
- Feedback ID: b92ba4de-bb85-4416-ab66-99896854666d
Errors
[]