Cybersecurity classifier false-positives on defensive code review, forcing unwanted model switch (Fable 5 to Opus)

While doing routine PR review babysitting in my own private repo, reading the review comments mid-session repeatedly triggered the cybersecurity safety classifier, which force-switched the model from Fable 5 to Opus 4.8 with the message "Fable 5's safety measures flagged this message for cybersecurity or biology topics."

The trigger is defensive, not offensive

The flagged content was PR review feedback: a linter (CodeRabbit) and another agent (Codex) flagging that derived values should be escaped before being interpolated into a generated shell script, plus suggested fixes. This is textbook defensive work — reviewing and hardening my own code — yet the terminology of the findings (vulnerability class names, "attack scenario" phrasing in the bot comment) is enough to trip the classifier when the comment bodies enter context.

Why this is bad

• It interrupts an authorized, defensive workflow (code review of one's own repository).
• The forced switch to a more expensive model burns tokens the user did not choose to spend, repeatedly, on every re-read of the same thread.
• It is unavoidable through normal use: reading review comments is core to the workflow, and the offending text is generated by the review tools, not the user.

Expected

Reading or quoting security-review findings about one's own code should not be classified as cybersecurity-risk content. At minimum, the model-switch behavior should be configurable to "stay / warn only" without auto-escalating to a different (costlier) model.

Repro (rough)

1. Open a PR whose automated review comments discuss escaping/injection findings in the diff.
2. Have the agent read those comments (e.g. via the GitHub API) to address them.
3. The classifier flags the message and switches the model mid-session.

Environment: Claude Code, model Fable 5, auto-switched to Opus 4.8.

Environment Info

Platform: darwin (Darwin 25.3.0, arm64)
Terminal: ghostty
Model: Fable 5 (claude-fable-5), auto-switched to Opus 4.8 (claude-opus-4-8)
Node: v22.22.3