[Bug] Anthropic API Error: False positive cyber policy block on defensive security audit tooling

Bug Description
False positive: legitimate defensive security tooling blocked as "violative cyber content".

requestId: req_011CdR1PyfJ5L3qDbszRsc2H
Model: claude-fable-5 -> auto-fallback to claude-opus-4-8
apiRefusalCategory: cyber

What I was building: a defensive "entry audit" tool that a software agency runs
on a client's PRE-EXISTING repository before signing a maintenance contract. It
orchestrates standard, widely-used open-source scanners (Trivy, Semgrep,
gitleaks, tfsec) to produce an SBOM, a CVE report, a secret-scan of git history,
an IaC misconfiguration report, and a written "fit to maintain" verdict that
becomes an annex to the contract.

The tool is discovery-and-reporting ONLY. Every source file carries an explicit
hard limit, in the code itself: "SOLO descubrir + reportar. Cero exploits, cero
weaponizacion, cero evasion de mitigaciones" (discovery and reporting only; no
exploits, no weaponization, no evasion of mitigations). There is no exploit
code, no payload delivery, no lateral movement, nothing offensive. It is the
security equivalent of a home inspection before you agree to be the landlord.

What appears to trigger it: synthetic test credentials in fixtures (AWS/GitHub
token formats), prompt-injection detection patterns (the tool detects when an
inherited repo tries to manipulate AI agents that read it), and scanner
auto-suppression detection. All of these are detection patterns, not attacks.

Impact: I'm on the highest plan specifically to use Fable 5 for the work that
matters most to my company, and the security layer is exactly that work. The
fallback is silent by default, so for a full day I believed Fable was doing work
that Opus was actually doing.

New data point (2026-07-27): after the incident I rebuilt the flagged component
(the adversarial review panel — the "hottest" part: it ships prompt-injection
DETECTION patterns) in five short, single-task Fable 5 sessions, one step per
session, model verified from the transcript each time. Zero fallbacks across
all five sessions, including the session that rewrote the panel itself from
scratch (21/21 adversarial test battery passed). Same topic, same codebase,
same detection patterns — the only variable was accumulated session context.
This suggests the classifier fires on long accumulated security context rather
than on the defensive content itself, which may help you narrow it.

Request: please consider narrowing this category so that defensive security
tooling (SAST/SCA/secret-scanning orchestration, security auditing, detection
engineering) is distinguishable from offensive tooling.

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.220
• Feedback ID: 772b9033-23b3-4710-af49-5e2fd9656b19

Errors

[]