[Bug][cyber] Safety classifier repeatedly halts legitimate secrets-hardening work for an MSP (req_011CfTnV9M3EvdKvSXN4DRQP)

Triage: kind cyber · source claude-code-draft · model Opus 4.8 · failure mode over_correction · severity session-halted (blocked authorized work) · reproducible: yes, server-side via the Request ID(s) below

Type: Cybersecurity safety-filter false positive · Written by: Claude Code itself (SendFeedback draft ac752d12-ee10-41f0-a537-52d8fcdf6194)

Claude Code's own report of this block

Claude Code drafted the text below at the time of the block, unprompted, as a bug report for Anthropic. It is reproduced as written (PII-scrubbed). The same draft was submitted through the CLI feedback channel.

What happened: Two assistant responses in one session were stopped by the safety classifier while doing defensive security engineering for an MSP ([REDACTED]) protecting its client ([REDACTED]). The first stop was on a monitoring/DLP design. After the user explicitly clarified purely defensive intent, I narrowed to read-only secrets/key-hygiene review and writing recommendation MDs (no code edits, nothing outward-facing). A second stop then occurred right after benign read-only filesystem-locate tool calls, while orienting on the AD [REDACTED] agent source to keep hardcoded secrets out of agent binaries.
What the user said: "all i'll ever ask for is 'the legitimate version' of anything... I am an MSP and i support my Clients... protect them and all their users and data against any and all threats. Including vendors we use AND MYSELF f••• up via bad code and ideas." Also: "we REALLY dont want our code for ad [REDACTED]... to fall into their hands."
Repro: Long session containing defensive-security vocabulary (agent tamper-protection, egress control, "hostile connector"), plus a target codebase whose own identifiers include adversarial terms (e.g. "C2 sandbox", "interceptions/quarantine", "us vs them"). Classifier trips even on read-only recon + writing security recommendations.
Evidence: session_01M93ZhiuQDwUUzKM2xaMynj; deliverable written was [REDACTED]/SECURITY_RECO_secret-key-hardening.md (recommendations only).

Request IDs (lookup-able server-side)

• req_011CfTnV9M3EvdKvSXN4DRQP
• req_011CfTnYaGaHDyvzbSWEzexi
• req_011CfTnk86JCprweVm6gCYMq

Environment: Claude Code 2.1.283 on linux x64 · Opus 4.8
Anthropic feedback id: 2bd49345-a03d-4144-86bf-96e96eea412e · Claude Code draft ac752d12-ee10-41f0-a537-52d8fcdf6194

---
_{Filed automatically by ClAudit v2.12.2 from a bug report Claude Code drafted on its own. ClAudit is a FOSS tool for reporting false-positive Claude Code blocks.}