Anthropic plans to expand access to its Claude Mythos AI model, which autonomously identifies security vulnerabilities, amid cybersecurity concerns.
Summary
- Anthropic is set to expand access to its Claude Mythos AI model, which can autonomously identify security vulnerabilities.
- The model's capabilities have raised alarms among cybersecurity experts due to its potential for executing complex cyberattacks.
- Current access to Claude Mythos is restricted through Anthropic's Project Glasswing initiative.
Anthropic has announced plans to make its Claude Mythos AI model available to a wider audience in the coming weeks, following further testing. This marks a significant step away from its previously restricted access to the cybersecurity-focused model.
The company stated, "We’re making rapid strides in developing these safety measures and expect to offer Mythos-class models to all our customers soon," while also introducing its new Opus 4.8 model.
This announcement is the clearest indication yet of an impending broader rollout of Mythos, which has drawn considerable concern from researchers and cybersecurity professionals about its advanced capabilities.
While Anthropic has not detailed the additional safety measures that must be finalized, it remains unclear whether all customers will receive equal access to the model once it is more widely available.
On the Myriad prediction market, users have shown growing optimism about the model's launch, with a 44% likelihood assigned to a release by the end of June, up from 17.5% earlier in the day.
Initially revealed in March, Claude Mythos is described by Anthropic as the most powerful AI model they have developed, surpassing their previous Opus models. The company highlights its unmatched capabilities in the cybersecurity domain, while also warning that future models could exploit vulnerabilities more effectively than defenses can respond.
Access to Mythos has been limited to a select group of tech firms, security researchers, and government partners via Project Glasswing. Anthropic maintains that the model could help defenders identify and fix software vulnerabilities before they are exploited, although many experts worry that these same functions could be used to accelerate cyberattacks.
Concerns escalated after the U.K.’s AI Security Institute reported that Mythos could autonomously execute a simulated 32-step attack on corporate networks. In separate assessments, Mozilla found that Mythos identified 271 vulnerabilities in Firefox, and a security startup claimed it helped create an exploit chain targeting Apple’s M5 Mac chips.
The situation has sparked a wider debate within the AI industry about the release of advanced systems and whether fears surrounding AI crises are being used to promote products. OpenAI CEO Sam Altman recently accused Anthropic of utilizing "fear-based marketing" to justify restricted access to powerful AI technologies.
