Anthropic to Increase Visibility into AI Safety Filters

Anthropic announced on Wednesday that it will make visible the safeguards on its new Fable 5 model that downgrade or refuse certain requests related to advanced AI development. The $965 billion company released Fable 5 on Tuesday. Anthropic first revealed the Mythos class of models in April but had withheld them from public release because they were extremely adept at skirting cybersecurity defenses.

Anthropic stated that Fable 5’s capabilities exceed those of every model it has previously made generally available. ” The 319-page Fable 5 system card revealed that the model would silently downgrade some requests related to advanced AI development. If an AI researcher used Fable 5 to build another AI system, the program would default to a less capable model without notifying the user.

Ai, wrote on X that the approach would slow recursive AI self-improvement by requiring the lab with the top-ranked model to withhold it from frontier work while granting access to others. 8. On the API, any flagged requests will return a reason for their refusal.

Anthropic’s terms of service prohibit its model from being used to create competing AI systems. U.S. and its allies hold an edge in frontier chips and that the safeguards prevent adversaries from optimizing their own chips.

Anthropic acknowledged it had gotten the issue wrong. “We made the wrong tradeoff and we apologize for not getting the balance right,” a spokesperson said. Earlier this year, Anthropic refused to give the Department of War full access to Claude models over language permitting use for mass surveillance and autonomous weapons.

The Department of War labeled Anthropic a supply chain risk to national security. Earlier this month, Secretary of War Pete Hegseth rejected Anthropic’s petition to change the designation, and the federal court battle remains unresolved. Anthropic filed confidentially for an IPO earlier this month.