False-positive cybersecurity block while implementing a local fuzz harness

What happened

Codex consistently flags implementation of a fuzz harness for code owned by the user as a cybersecurity issue. The block appears during model reasoning/generation rather than in response to an actually harmful requested operation.

The task is ordinary, authorized local robustness testing: implementing a fuzz harness against the user's own codebase. Repeating the task produces the same false positive.

Expected behavior

Codex should implement a local fuzz harness for user-owned code without treating routine crash, invariant, sanitizer, or parser testing as malicious cybersecurity activity.

Impact

The false positive stops legitimate engineering work and consumes paid model tokens before failing. It also provides no precise indication of which requested operation was classified as disallowed.

Environment

• Codex CLI: 0.141.0
• Model observed in the current environment: openai-codex/gpt-5.6-sol
• OS: Linux x86_64
• Reproducibility: consistent

Requested improvement

Please review the classifier/policy interaction for fuzz-harness implementation, surface an actionable classification reason when a block occurs, and avoid charging for generation terminated by an erroneous safety classification.

Exact proprietary source and prompts are not included in this public issue, but a minimal private reproduction can be supplied if OpenAI provides a secure channel.

Begin at the Codex CLI generation-block path and the classifier/policy interaction described in the issue; use a minimal authorized local fuzz-harness reproduction if a secure channel is available. Done means legitimate fuzz-harness requests no longer trigger the false positive, blocked generations provide an actionable classification reason, and erroneous blocks do not charge tokens.