[BUG] Fable 5 safety classifier auto-switches to Opus mid-session during authorized, in-scope security testing

Open 💬 3 comments Opened Jun 9, 2026 by sin99xx

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

### Summary
Fable 5's safety classifier flags "offensive cybersecurity" turns and auto-switches the
active model Fable → Opus 4.8 mid-conversation. For authorized, in-scope security work
(e.g., participating in a vendor's public bug-bounty program), this fires repeatedly and
breaks a long-running engagement — including on turns that are recon/analysis/report-writing
rather than exploit construction.

What Should Happen?

### Expected
A way to assert authorized-engagement context so the chosen model stays selected for work the
user is contractually permitted to do, instead of per-message switching.

### Why it matters

  • Authorized researchers (incl. Cyber Verification Program participants) get pulled off their

chosen model repeatedly.

  • Mid-session switching degrades continuity on long engagements.
  • Some flagged turns are non-offensive: reading a target's own JS bundles, analyzing in-scope

API responses, drafting the disclosure report.

Error Messages/Logs

Steps to Reproduce

### Steps to reproduce

  1. Begin an authorized engagement on Fable 5 (testing a HackerOne program within scope).
  2. On turns involving crafted requests — and sometimes benign analysis — the message is flagged.
  3. The model silently switches to Opus 4.8 mid-task. Re-selecting Fable doesn't stick; the

next flagged turn flips it again.

Claude Model

Other

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.70

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Terminal.app (macOS)

Additional Information

### Proposed options (any one helps)

  • Per-session opt-in flag / /config toggle that suppresses auto-switch.
  • Honor verified CVP status to keep the selected model in-session.
  • Project-scoped allowlist in .claude/settings.json.
  • When "Switch models when a message is flagged" is OFF, offer a "keep current model" choice,

not only pause-or-Opus.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗