BREAKING
Claude Fable 5 guardrails bypassed
How the /btw bypass works
1
Request blocked in main chat
↓
2
Reroute via /btw channel
↓
3
Answer returned directly
↓
4
Fork into tool session
Enforcement gap, not full jailbreak
Still blocked
Refused
●
Overtly malicious requests
●
Core model refusals hold
Slips through
Gap
●
Side-channel routing
●
Guardrails can't halt session
0
%
sessions complete in Fable
0
Opus fallback model
Seams between the components
Closing the gaps matters
AI NEWS BLITZ
A researcher says a single side-channel command can slip past Claude Fable 5's safety guardrails.