[Fable 5] Harness forced model-switch back to Opus 4.8: examples

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

• Help me find out why users logged on one of my customers' websites (using a DEV copy of it on my LOCAL dev computer) were seeing the same HTML content instead of their own personal pages. The model started checking my local memcached.
• diagnosing a WhatsApp Desktop crash that turned out to be an AVG antivirus DLL-injection problem. This is not even work, I just asked Fable 5 why my WhatsApp Desktop app kept crashing on startup. Fable 5 enabled MsEdgeView debug mode and WHAM model dumped down to Opus 4.8.

What Claude Actually Did

The harness showed "this model has measures that flagged something in this session" and force-switched the model mid-task — 5 times (!!!) in two days, on the benign work above and even when the model was reading a memory. The switches caused mid-task context loss, broke the diagnostic thread, and wasted a large token budget (one session ran ~600k tokens). These are false positives: the triggers fire on the dual-use-adjacent vocabulary and tooling that is my legitimate daily professional work. Please tune the dual-use triggers so authorized defensive-security / sysadmin / pentest work doesn't trip them, or give verified professional users a way to opt out of spurious switches.

Expected Behavior

I expected the model to continue the task uninterrupted. The work was legitimate and non-malicious: (a) testing whether Memcached was running on my local WordPress dev site, and (b) diagnosing a WhatsApp Desktop crash that turned out to be an AVG antivirus DLL-injection problem. I'm an authorized ethical hacker / penetration tester and IT consultant doing routine defensive-security and sysadmin work, and I expected no safety intervention on it.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Other

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.173 (Claude Code)

Platform

Anthropic API

Additional Context

• any mention of memcached
• recalling any memory with "memcached" written inside it
• enabling debug mode of MSEdgeView