Summary
- Anthropic anticipates making its Claude Mythos AI model available to customers within the next few weeks after completing further testing.
- The model has raised alarms as researchers discovered it can autonomously identify security flaws and execute complex cyberattacks.
- Access to Claude Mythos is currently limited via Anthropic's Project Glasswing program.
On Thursday, Anthropic announced that it plans to broaden access to its Claude Mythos AI models “in the coming weeks,” suggesting that the company is preparing to transition from the tightly controlled release of this cybersecurity-oriented system.
“We’re making rapid strides in developing these safety measures and expect to offer Mythos-class models to all our customers soon,” Anthropic stated while unveiling its new Opus 4.8 model.
This announcement marks Anthropic's most definitive indication that a more extensive rollout of Mythos is forthcoming, following numerous alerts from researchers, governmental bodies, and cybersecurity professionals regarding the model's capabilities.
The company did not specify what additional safety measures need to be finalized before a wider rollout of Mythos or whether all customers will have equal access once it becomes more generally available.
Anthropic has not responded to a request for comment from Decrypt.
Users on Myriad—a prediction market platform run by Decrypt's parent company, Dastan—are increasingly optimistic that the model will launch by the end of June, assigning a 44% likelihood to that scenario, a significant rise from 17.5% earlier in the day.
Claude Mythos was first introduced in March following the leak of draft materials from Anthropic's blog, where the company described it as “the most powerful AI model we’ve ever developed,” categorizing it as a new level above its advanced Opus models.
“While Mythos is currently unmatched in cyber capabilities, it signals the arrival of future models that can exploit vulnerabilities far more effectively than defenders can respond,” the company noted.
Since then, Anthropic has limited access to Mythos through Project Glasswing, which provides a select group of tech firms, security researchers, and government partners controlled access to the model.
The firm contends that Mythos could assist defenders in identifying and rectifying software vulnerabilities prior to their exploitation by attackers. However, concerns have been raised by security researchers and government entities that these same functions could expedite cyberattacks.
These worries intensified when the U.K.’s AI Security Institute revealed that Mythos could autonomously conduct a 32-step simulated corporate network attack during testing. In April, Mozilla reported that Mythos detected 271 vulnerabilities in Firefox during its internal security assessments. Earlier this month, security startup Calif claimed that a preview version of the Mythos model aided researchers in creating an exploit chain targeting Apple’s M5 Mac chips.
The model has also ignited a broader conversation within the AI sector regarding the release of advanced systems and whether the fear of an AI crisis is being leveraged to market products.
In an interview last month on the “Core Memory” podcast, OpenAI CEO Sam Altman accused Anthropic of employing “fear-based marketing,” suggesting that cybersecurity risk warnings could serve as a rationale for restricting access to powerful AI technologies.