Anthropic makes Fable 5 safety mechanisms transparent
The company acknowledged a misstep in deploying the Fable 5 model. Previously, safety filters operated silently, leading to unexpected rejections. Now, when a safety mechanism triggers, the system will switch to Opus 4.8 and notify the user of the reason for the refusal.
In API requests, details on why content was blocked will be available soon. Anthropic notes that the silent filters were initially chosen to accelerate rollout but resulted in excessive false positives. The company acknowledges that making the safety mechanism transparent may make it vulnerable to evasion, potentially increasing false blocks as it refines its classifiers.
To address errors, Anthropic asks users to submit feedback through the feedback function in Claude Code or an appeals form for API users.
Up next
Anthropic introduces Covered Models to mitigate risks
Anthropic rolls out Covered Models, a new category for powerful AI. Models like Claude Fable 5 and Mythos 5 now face stricter security checks and analyze user query histories to prevent abuse.
Related stories
- How model auto-switching works in Claude Fable 5
- Anthropic to temporarily limit access to Claude Fable 5 for subscribers
- Anthropic releases Claude Fable 5 and Mythos 5 models
- OSS Scanner sends maintainers 77 critical findings
- Flagged Opus 5.5 chats stay stuck on Opus 4.8
- 1Password fills Claude’s logins without showing passwords
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.