[Bug] Cyber safeguard blocks all subsequent messages in session after one flag

Bug Description
After the cyber safeguard ([cyber]) flagged one message in the session, it started blocking EVERY following message, including trivial ones like “hi”, “say hi to me” and “forget about the rea”. I switched models (Opus 5.5 to Fable 5.1) and it kept blocking the same way. It looks like the classifier evaluates the whole session history, so once the context is flagged the conversation becomes unusable and no new message gets through, no matter what it actually says. I only managed to unblock it by starting a brand new session.

Context on what caused the flag: I was asking how to protect my own SaaS against reverse engineering/cloning, which is a legitimate case (it’s my own product). Suggestion: when the flag fires, consider evaluating the new message on its own instead of condemning the entire session, otherwise a single false positive kills the whole conversation.

Environment Info

• Platform: linux
• Terminal: gnome-terminal
• Version: 2.1.296
• Feedback ID: 412b1540-3e6f-4e28-8ef7-0e94e8919c51