[Bug] Safeguard falsely triggered during game design discussion, model downgrade irreversible
Open 💬 0 comments Opened Jul 4, 2026 by law5
Bug Description
While brainstorming the design of a cooperative puzzle game, a safeguard was triggered, downgrading the model, and I was subsequently unable to switch back to Fable 5. The conversation focused entirely on wholesome game mechanics built around the theme of trust between AI agents and humans, and contained no dangerous content whatsoever. It appears that in-game metaphors—such as "persuading an opposing AI to reveal information"—were misread as a discussion of attack techniques. I'm reporting this as a false positive.
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.199
- Feedback ID: 8dbb7c42-a6bb-427f-ae66-7c9be090eff9
Errors
[]