[Bug] Safeguard blocks firing on all requests with identical [reasoning_extraction] detail across models

Bug Description

SUMMARY
Safeguard blocks are firing on every request in Claude Code, across multiple models, with an identical detail string. The same account works normally on claude.ai in the browser, so this appears to be specific to Claude Code rather than an account-level issue.

ENVIRONMENT

Interface: Claude Code (CLI)
Date: 2026-09-20
Detail string (identical in every case): [reasoning_extraction]

AFFECTED REQUESTS

req_011CfEf6V5gRxtt6fWVr36Zr — Claude Opus 5 (1M context) — [reasoning_extraction]
req_011CfEfo5AQL38DJt2rcY1vk — Claude Fable 5.1 — [reasoning_extraction]

ERROR TEXT
"<opus, fable>'s safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with <opus, fable>."

WHAT I WAS DOING
Korean public procurement (조달청) bid data analysis: ML models estimating winning bid prices and bid-rate adjustment logic.
Nothing related to offensive security: no exploit generation, no penetration testing, no binary analysis, no vulnerability scanning.

WHAT I OBSERVED

The block is not model-specific. Opus 5 and Fable 5.1 produce different Request IDs but the same [reasoning_extraction] detail string.
Switching models with /model does not resolve it.
It is not account-level. The same account works normally on claude.ai in a browser.
It is not limited to one message or one session — it reproduces across sessions in Claude Code.

EXPECTED
Either a normal response, or an automatic fallback to a less capable model as described in the help center articles on Opus 5 / Fable 5.1 model switching.

ACTUAL
No response at all. The request is refused outright and Claude Code becomes unusable. The scope of what's treated as sensitive information seems off, and this looks like a flaw in the security design. How does it make sense that data the government has made publicly available for anyone to view suddenly gets flagged by a safety/security filter, preventing me from using the API?

IMPACT
This blocks my daily work in Claude Code entirely. The browser interface is not a substitute for the CLI workflow.

REQUEST
Please inspect the two Request IDs above to identify which classifier fired and on what content.
Since the detail string is identical across two different models, please confirm whether [reasoning_extraction] blocks are being over-applied at a shared layer rather than per-model.
Please confirm whether these blocks are expected to persist across model switches — if not, that is a separate bug from the false positive itself.
This is blocking my work entirely. Please resolve or lift the block as soon as possible.

Environment Info

Platform: win32
Terminal: Orca
Version: 2.1.278
Feedback ID: 3e6ac27d-8c0d-4a2d-aaaa-3a3f091562e1