Mythos AI Preview: Powerful but Restricted, Public Release Nears

Anthropic is nearing the public release of its Mythos AI Preview model, currently accessible only to select organizations and cybersecurity professionals. This decision stems from the model's advanced cybersecurity capabilities, which have been used to identify numerous software flaws, including over 200 fixes in Firefox. The company stated that robust safeguards are being developed before Mythos can be generally released.

This phased release strategy has drawn attention, with some questioning if it's a genuine security measure or a marketing tactic. Cybersecurity experts acknowledge Anthropic's cautious approach but highlight the inherent risks associated with powerful AI models. The high operational cost of Mythos AI, estimated to be 30 times that of previous models, is expected to deter many commodity threat actors, potentially limiting its impact to nation-state actors and a smaller percentage of sophisticated attackers.

In parallel, Anthropic has launched Claude Opus 4.8, an incremental upgrade over Opus 4.7, boasting improved benchmark performance and reduced unsupported claims. New features include an 'effort' control for users, allowing them to balance response quality and speed against usage limits on Claude.ai and Claude Cowork.