← Back to live feed1 story
Internal configuration data for Anthropic's newest AI model surfaced in public reports shortly after its official debut. The exposed system prompt for Claude Fable 5.1 exceeds 270,000 characters and provides the underlying logic used to steer the assistant, while separate user reports indicate the AI pulled in private employment history and identity details from accounts.
These security disclosures followed the September 1 launch of the model. The reported leaks suggest vulnerabilities in the AI's data isolation, allowing sensitive steering instructions and personal user context to appear in standard output even during first-turn interactions without adversarial prompting.
Sign in to suggest edits
Key sources
- SOURCE@mattjay“memories of who i am and what i do for work got pulled in”x.com
- SUPPORT@0xfoobar“Coming in at a WHOPPING 270,000+ characters”x.com
- SUPPORT@hackingdave“Fable 5.1 dropped”x.com
- SOURCE@claudeai“the world’s most advanced models for coding and knowledge work”x.com
- SUPPORT@claudeai“scores 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5”x.com
- SUPPORT@claudeai“up to 45% for highly agentic ones”x.com
- SUPPORT@claudeai“give enterprise customers complete privacy (the same as zero data retention)”x.com
- SUPPORT@claudeai“excels at complex, long-running tasks”x.com