American startup Anthropic announced the implementation of security measures to restrict cybersecurity-related requests in its model initially made available in April only to a limited group of partners.
Anthropic began offering a limited, consumer-oriented version of Mythos, its artificial intelligence model whose access was controlled because of its strong cyberattack capabilities. Called “Claude Fable 5”, the new option represents “a Mythos-type model, but which we have made safe for general use”, informed the company responsible for the Claude assistant in a blog post this Tuesday, June 9.
“Without security measures, Fable 5’s cybersecurity features could be misused and cause serious damage,” highlighted Anthropic. “So we decided to release it publicly with safeguards that redirect queries on certain topics to our next most powerful model, Claude Opus 4.8.”
This limitation applies to user requests about cybersecurity, but also about biology and chemistry, according to Anthropic, which sees a risk that the models facilitate the creation of biological weapons. The company also cited “distillation”, a technique in which a smaller model consults a larger one to copy it — a practice that, according to the company, would have been used by Chinese agents. When detecting a request in these areas, Fable 5 should decline and forward the query to Opus 4.8.
