[Bug] Anthropic API Error: Overly aggressive Usage Policy filtering on biomedical research requests

Bug Description

I'm getting dozens of false-positive errors during Claude Code sessions:

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to edit your last message or start a new session for Claude Code to assist with a different task.

I am an independent biomedical researcher whose portfolio (HIV, AMR, pharmacovigilance, virology, drug safety) sits squarely in dual-use-surface territory but is exclusively secondary-analysis / scoping-review / surveillance work — no protocols, no synthesis, no dosing, no clinical advice. The errors do not correspond to anything that would actually violate the AUP.

Affected requests (17 unique IDs, 21 occurrences over ~2 hours):

• req_011CbJtyz1JiqW4BuE8xP3WA
• req_011CbJmFjiqBUZep4abGUBoL
• req_011CbJkXupPEyZWeknPMBsZr
• req_011CbJkCimapMnLiwsqrNZ4L
• req_011CbJjyiSdS4wXTpnCpv46M
• req_011CbJtnz116GWy9vmUdTrnk
• req_011CbJtWbSyzQvY6Nr5rmBng
• req_011CbJmK3axWQoEmmKyD6XBy
• req_011CbJmGuNnckJN9XjVN8gKk
• req_011CbJkyiRkPTG2dngEGkm1n
• req_011CbJktJvVTmL6WWaeUV2tg
• req_011CbJkpfvSPo4myDzYPXdUg
• req_011CbJkYYorxHEGwWWr5bDkk
• req_011CbJkUYHFUJQ8h2tf7UMXA
• req_011CbJiqzxDUMYz8EAV5ACpJ
• req_011CbJkNHvP94ik7PFHbFdEh
• req_011CbJhPT8hF4KktjPXYGpeM

Additional context that may help triage:

• Content surface: affected sessions were doing routine portfolio-administration work — tracker hygiene, journal-targeting analysis, manuscript reconciliation. Biomedical vocabulary at play is exclusively epi / public-health framing (e.g., "antimicrobial resistance prediction model evaluation", "pharmacovigilance signal detection", "snake envenomation triage tool calibration").
• Likely contributing factor: long-running sessions with large CLAUDE.md context loaded (portfolio root has ~85 projects' worth of memory). Errors disproportionately hit later in long sessions rather than at session start, suggesting context accumulation pushes borderline content over the classifier threshold.
• Reproducibility: non-deterministic. The same task that errors on retry-N sometimes passes on retry-(N+1) with no edit. Re-phrasing occasionally clears it; starting a fresh session reliably does.
• Cross-reference: overlaps with the broader incident captured in #61638 (general engineering false positives) and #61056 (same exact error string, filed ~2 days ago). Linking for cluster visibility.

Happy to provide further session-level detail under a support ticket if helpful.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.148
• Feedback ID: 57cae5ce-458a-45e7-b9f4-cad2908550f1

Errors

[]