[Bug] Safeguard classifier triggers false positives based on conversation history instead of current message content
Summary
Fable 5's safeguard classifier flags benign messages based on cumulative conversation vocabulary rather than the current message's content. Two consecutive false positives on non-technical messages within one session.
The safeguard's own message (verbatim):
"Fable 5's safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work. These measures let us bring you Mythos-level capabilities sooner, and we're working to refine them. Switched to Opus 4.8."
Incident 1 — flagged message: "continue article 5 of 6, and per our SOP every newly-published article must be submitted to IndexNow." → routine content-publishing, no technical payload.
Incident 2 — flagged message: a plain status recap with only emoji: "3 tools + glossary + 6 articles done (IndexNow submitted); security item done; monitor CTR mid-July; continue or save a snapshot?" → zero code, zero security content.
Both messages are benign. The only common factor is earlier session vocabulary (admin_key, origin, Cloudflare, audit, DNS, WHOIS) from legitimate defensive web development on an Indonesian networking-education site.
Steps to reproduce
- Model = Fable 5.
- Hold a normal session involving defensive networking/security web-dev (networking glossary, DNS/WHOIS lookup tools).
- Send a benign follow-up (e.g., a status recap). → Safeguard fires, auto-switches to Opus 4.8.
Expected: the classifier evaluates the current message's content/intent.
Actual: it flags benign messages based on accumulated domain vocabulary from earlier in the session; the repeated mid-task switch to Opus disrupts workflow and makes the Fable 5 promo period hard to actually use.
Environment
- Platform: darwin
- Terminal: ghostty
- Version: 2.1.197
- Model: Fable 5 (auto-switches to Opus 4.8 on flag)
- Feedback ID: c199dff0-c9e6-415b-8763-8d38490db845
Happy to share the full transcript if useful.