Safety classifier false-positives block legitimate automation on Fable 5 / Opus 5, with no fallback

What happened

Fable 5 and Opus 5 refuse legitimate business-automation requests via the safety classifier ("safeguards flagged this message"). The turn ends with the AUP error and no answer, and Claude Code provides no automatic fallback to another model.

Example flagged request (routine data-automation for my own workflow, paraphrased): "find a way to pull the MMR value from the same source we already use, without the login/password step." This trips the cyber classifier even though it's a legitimate scraping task on a source I already have access to.

Why this is a false positive / bug

• The model is reachable — light prompts on claude-fable-5 succeed (verified via claude -p --model claude-fable-5). So this is not an access, credits, or rate-limit issue. Subscription is max with extra usage enabled.
• Switching model doesn't help — moving from Opus 5 to Fable 5 hits the same (or stronger) classifier, so there is effectively no usable model for these tasks.
• No fallback / no recovery — the turn just dies. There's no one-key "retry this turn on another model"; the user has to double-esc and re-edit, which loses the turn.

Confirmed Request IDs

• req_011Cdb631RVtWAXoyLo4kTuc
• req_011Cdb6bjwkLhsdbZbM1Ss8B

(both 2026-08-01, model Opus 5; same classifier fires on Fable 5)

Exact error shown

API Error: Opus 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Claude Code can't respond to this message with Opus 5.

Ask

1. Reduce false positives on legitimate automation / scraping tasks.
2. Add an automatic fallback (or a one-key "retry this turn with another model") instead of dropping the turn.

Environment

• Claude Code: 2.1.220 (Claude Code)
• OS: Linux 6.8.0-90-generic
• Subscription: max (extra usage enabled)