Fable and Mythos 5.2: The System Card
What happened
Anthropic has launched Claude Mythos 5.2, a new AI model with enhanced safeguards, particularly in biology and cybersecurity. The system card for Mythos 5.2 outlines its technical abilities, alignment risks, and findings from model welfare assessments. The model shows moderate improvements over earlier versions, with no major surprises in welfare evaluation. Mythos 5.2 is a variant of the same underlying model as Fable 5.2, but with more permissive safety controls in specific domains.
Key details
- Claude Mythos 5.2 was released with updated biology and cybersecurity safeguards.
- The system card includes assessments of alignment risk and model welfare.
- Mythos 5.2 is moderately more capable than previous versions, according to evaluations.
- Fable 5.2 and Mythos 5.2 are the same model, but Mythos has more permissive safeguards.
- A recent report estimated Anthropic's internal AI R&D progress could be accelerated by up to 1.9x using a similar model.
Why it matters
Organizations using Anthropic's models in sensitive fields like biology and cybersecurity now have access to a version with updated safety measures and slightly higher capabilities. Companies and researchers relying on these models may see improved performance, but must also consider the new alignment risk assessments and operational issues noted in the system card. Anthropic itself faces the need to monitor and address the increased pace of internal AI research and development enabled by these models.