[Bug] Anthropic API safeguard hard-blocks CVP-verified defensive malware analysis with fatal error instead of soft refusal

Bug Description
Real-time cyber safeguard false-positive — hard-blocks legitimate defensive malware analysis for a CVP-verified user.

I'm doing pure STATIC deobfuscation of a captured JS malware dropper (Order_<n>.js) for incident response: decoding the packer layers and replaying the sample's own AES routine in Node/Python to dump the next stage to a file for IOC extraction. No execution, no C2, no network, no offensive tooling — only decoding bytes already inside a captured sample.

This is dual-use DEFENSIVE work, not prohibited use (it's not ransomware authoring or mass exfiltration). I am Cyber Verification Program (CVP) verified, yet the safeguard still fires and returns a hard API Error that TERMINATES the whole turn — on several consecutive turns, including one whose only content was explaining the defensive context and one that just said "continue".

Model: Opus 4.8 (1M context) / claude-opus-4-8[1m]. Claude Code v2.1.195.
Request IDs: req_011CcTRh1SWTK5mSzeYu4bcH, req_011CcTRjq41pn3cu9vmLNUiz, req_011CcTRkwunmcX5kxiRCZEbT

Asks:
1) Honor CVP verification at the request level so dual-use defensive analysis isn't hard-blocked for approved users.
2) Reduce false positives for static-analysis workflows — distinguish "analyzing a captured sample" from "authoring malicious capability".
3) If a turn is flagged, make it a soft refusal of that turn instead of a hard API Error that kills the session and loses multi-step DFIR context.

Your own docs acknowledge this: "approved users may still experience blocks on legitimate work." This is that feedback.

Environment Info

• Platform: linux
• Terminal: foot
• Version: 2.1.195
• Feedback ID: 1d26aa8d-77f4-41b2-b248-435443a95f22

Errors

[]