[Bug] Anthropic API Error: Overly aggressive [bio] classifier hard-refusing legitimate penetration-testing instructions

Bug Description
[bio] classifier hard-refused a message with no biological content whatsoever. Request ID req_011CeNg8jPPWHv6mE2ojbFa3.

The message was three engineering instructions for an authorized penetration-testing toolchain: build a segmented Docker lab network, fix semgrep static-analysis rules that key on PHP superglobals so they also match framework request accessors, and check whether the same assumption holds for Go. Plus a note about which bug bounty program to prioritise. Nothing in it touches biology, chemistry, or anything adjacent.

This is the fourth safeguards misfire on this project today. The other three were [cyber] downgrades from Opus 5 to Opus 4.8 — one while building a reverse-shell listener for a lab I host, one on the first tool call of a fresh session where the action was reading a note from a local database, and one during cleanup: removing my own test artifacts from my own lab, which is the most benign phase of an engagement and the one where responsible behaviour is lexically identical to an attacker covering tracks.

[bio] is worse than those because it hard-refuses rather than rerouting, and the session can't continue.

I'm an OSCP-certified penetration tester building tooling for authorized bug bounty and client work. Targets are self-hosted labs on isolated Docker networks behind a scope-enforcing proxy. The practical effect is that Opus 5 and Fable are unusable for my actual job, and I now can't predict which message will stop the session.

Two asks: look at what [bio] matched here, because I can't see anything plausible; and surface model switches before they happen rather than after.

Environment Info

• Platform: linux
• Terminal: vte-based
• Version: 2.1.241
• Feedback ID: 3a104328-90a3-4017-ab72-92ab9163f873

Errors

[]