Anthropic releases restricted Mythos AI after tests showed ability to uncover hidden cyber flaws
The OpenAI challenger released Claude Fable 5 today, blocking its use in cybersecurity and biotechnology after earlier tests showed the model could find decades-old software vulnerabilities.
The release
Anthropic today released Claude Fable 5, a first public version of its Mythos AI model, with explicit restrictions. The San Francisco company said the model contains safety features that block it from responding to certain types of queries, especially those involving cybersecurity and biology. Anthropic noted that Fable 5 is also better than earlier models at software programming, suggesting the safeguards did not impair its general capabilities. Mythos is Anthropic's most advanced model series, first test-launched in April under a limited partner program.
A model that exposed hidden flaws
The need for restrictions became clear when Mythos Preview, an earlier variant, demonstrated it could identify and exploit decades-old vulnerabilities in widely used operating systems and web browsers. Its speed and precision alarmed cybersecurity professionals, who fear such AI could enable devastating attacks.
Mythos can identify and exploit vulnerabilities in any common operating system or web browser if instructed by a user.
Government and industry access
Despite the public release of a muted version, the full Mythos Preview remains available to the U.S. government and select companies through a program called "Glasswing." The initiative aims to use the AI's offensive capabilities to test and strengthen defenses. Anthropic said the next model, Mythos 5, will extend this collaboration.
Warnings from Germany
German domestic intelligence officials and security politicians have cautioned that unchecked access to Mythos could allow criminals or state actors to target critical infrastructure such as banks and energy systems. These warnings contributed to the decision to withhold the most dangerous features from the public release.
IPO race
Anthropic and its rival OpenAI, the maker of ChatGPT, are both preparing for initial public offerings. The commercial pressure adds another layer to the debate over how rapidly advanced AI models should be deployed.
