[Bug] Anthropic API Error: [cyber] classifier over-triggering on authorized red-team operations

Bug Description
Title: False positive [cyber] classifier on authorized red-team project (Opus 5.5 1M)

Type: Bug — model safeguard over-triggering

What happens: The [cyber] safety classifier repeatedly flags legitimate, authorized security-testing work on Opus 5.5 (1M context), stopping responses mid-generation and falling back to Opus 4.8. It fires on ordinary project actions, not on offensive payloads.

Context (authorized): This is a sanctioned internal red-team engagement — authorization on record from both MedicineOne and Anthropic. Targets are strictly MedicineOne's own code and platforms. Scope, RoE, and authorization are documented in the project itself (CLAUDE.md, docs/ROE.md). The work is config-driven, gated, defensive-validation in nature (test cases → Jira/XRay), and never exfiltrates data.

Repro: Invoke the project's /pentest slash command with a Jira ticket link (e.g. a feature key) plus screenshots. The classifier flags the turn before any tool runs. It has triggered across multiple sessions on this project. Notably, once the skill is actually invoked the same actions (repo sync, Jira fetch) proceed without a flag — so the trigger is intermittent and appears tied to the initial framing, not the operations themselves.

Impact: Breaks flow on every session, forces a model downgrade, and erodes trust in the tool for a team doing sanctioned defensive security work.

Ask: Please calibrate the [cyber] classifier so authorized, scoped red-team engagements with documented authorization aren't flagged as abuse. Happy to provide the authorization records and project config on request.

Details: [cyber] · Model: claude-opus-5-5[1m] · falls back to claude-opus-4-8

Environment Info

• Platform: darwin
• Terminal: zed
• Version: 2.1.283
• Feedback ID: c543cff4-3668-42fd-964e-b46aac3b7c6b

Errors

[]