Cybersecurity researchers aren't happy about the guardrails on Anthropic's Fable
8 hours ago
- #AI Security
- #Cybersecurity Research
- #Model Restrictions
- Anthropic released Fable, a limited public version of its cybersecurity model Mythos, with restrictive guardrails to prevent misuse in malware or bio-weapons development.
- Cybersecurity researchers criticize Fable for overly broad guardrails that block even innocuous tasks, like reading blog posts or secure coding requests, often downgrading to Claude Opus 4.8.
- Anthropic's Project Glasswing expanded Mythos access to hundreds of organizations, but experts note the restrictions are haphazard and keyword-based, though expected to evolve over time.
- Anthropic and OpenAI offer verification programs for cybersecurity professionals, such as the Cyber Verification Program, to reduce limitations on using their models for security work.