Anthropic stated that the Claude Mythos model, designed for cybersecurity scenarios, is expected to become available to more customers in the coming weeks. This is the company's most definitive statement to date, indicating that the previously strictly limited model is nearing a wider release.

Currently, Anthropic has not specified what security measures still need to be implemented, nor has it disclosed whether different customers will have the same access after the model is made available. The company released this update along with the new Opus 4.8 model.

Previously limited to controlled use

Claude Mythos first came to the public eye in March of this year. At that time, a leaked draft of the Anthropic blog described it as the company's most powerful AI model to date and positioned it above the Opus series.

Subsequently, Anthropic restricted access to Mythos through Project Glasswing, granting it only access to select technology companies, security researchers, and government partners, and requiring its use under controlled conditions.

Security tests draw attention

The focus of external attention is that this model can not only help defenders discover and patch vulnerabilities, but may also accelerate the speed at which attackers launch cyberattacks.

The UK's AI Security Institute found in tests that Mythos could autonomously complete 32 steps of a simulated corporate cyberattack. Mozilla stated in April that the model identified 271 vulnerabilities in Firefox during an internal security assessment. Security startup Calif claimed earlier this month that a preview version of Mythos helped researchers build an exploit chain targeting Apple's M5 Mac chip.

Industry debate continues to intensify

As Mythos nears wider openness, the debate within the AI industry regarding how advanced models should be released is intensifying. One side argues that such models should be gradually released under stricter restrictions; another questions whether vendors are using concerns about high-risk AI to control access.

In a podcast last month, OpenAI CEO Sam Altman singled out Anthropic, saying that its statements on cybersecurity risks had the flavor of "fear-based marketing" and could be used to justify limiting the use of powerful models.