Eigen RadarAI
Archive
Analysis

Anthropic releases Fable and Mythos with different access safeguards

Anthropic released Claude Fable and Claude Mythos on November 9, using the same underlying model with different access arrangements. Fable is broadly available with additional protections against dangerous uses, while Mythos is restricted to vetted organizations. The company also described behavioral testing and external evaluations alongside the release.

Artificial Intelligence··Morning
An evaluator reads printed transcript pages by a window in an AI evaluation office.

One model receives different access controls

Anthropic, the developer of the Claude artificial intelligence models, released Fable 5.2 and Mythos 5.2 on November 9. Both use the same underlying model, according to the company. Fable is broadly available, while Mythos is limited to programs serving vetted organizations. Anthropic says it has restricted some capabilities as part of the release.[1], [2]

Fable has additional protections against cyber and biological misuse. Certain requests in these areas can be routed to earlier models. Cybersecurity professionals can apply for verified access to Mythos. These arrangements give the two products different safeguards and eligibility requirements despite their shared model.[1]

Behavioral audits include difficult sandbox tasks

Anthropic describes an automated behavioral audit involving roughly 1,900 scenarios, alongside a separate sandbox evaluation. That environment deliberately includes tasks mostly impossible to complete within its intended boundaries. The company says the models expressed suspicion that they were being evaluated more often than some recent predecessors. Internal transcript inspection, training monitors and external evaluations informed its assessment.[1]

Outside evaluators have a separate role

The UK AI Security Institute, METR and Frontier Design are named as external evaluators. Anthropic says they did not review its announcement or system card and do not endorse its conclusions. A service running within customer infrastructure is also entering private beta, with Bridgewater, Salesforce and Stripe named as customers. The release includes separate access controls and data-retention arrangements.[1]

References

  1. News sourceAnthropicAnthropic releases Fable and Mythos 5.2↩1↩2↩3↩4
  2. News sourceClubicAnthropic releases Claude Fable 5.2 and Claude Mythos 5.2↩
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →