Summary
- Anthropic is broadening the reach of Project Glasswing, allowing additional organizations to utilize its advanced Claude Mythos AI model.
- The firm recently indicated that wider customer access to Mythos-class models could be available in the next few weeks.
- Researchers and government entities continue to examine the model's capacity to detect vulnerabilities and perform intricate cyber operations.
Anthropic is widening the availability of its Claude Mythos AI model via Project Glasswing, a program designed for technology and security companies, as well as governmental bodies, to identify and rectify bugs and potential vulnerabilities prior to the public launch of this sophisticated model.
In a blog post released on Tuesday, Anthropic announced that it is incorporating approximately 150 new organizations into its controlled program as it gears up for a more extensive rollout of the contentious model.
This expansion follows Anthropic's statement that it anticipates making Mythos-class models available to clients "in the coming weeks" once additional safety measures are finalized. This announcement coincides with Anthropic's preparations to go public after it confidentially filed for an IPO on Monday with the U.S. Securities and Exchange Commission.
“After several weeks of intensive collaboration with our Project Glasswing partners, including the security sector, open-source software maintainers, and the U.S. government, we’re extending our partnership to around 150 new entities,” stated Anthropic. “Each organization must meet our security criteria before gaining access.”
The recent expansion of Project Glasswing includes organizations from essential infrastructure sectors like power, water, healthcare, communications, and hardware. Anthropic noted that several of these new partners manage software utilized by governments and hundreds of millions of users.
“What unites each partner is that a successful breach of their code could lead to severe consequences,” Anthropic explained. “For the majority of partners, we estimate that a significant attack could impact over 100 million individuals, posing serious implications for both global and national security.”
Claude Mythos was first publicly introduced in March after preliminary materials from Anthropic leaked online. According to Anthropic, organizations that previously gained access have already utilized Mythos Preview to discover over 10,000 high or critical-severity software vulnerabilities. These partners are also employing the model to draft patches, perform security assessments, and test software before its release.
The capabilities of the model have sparked concerns among researchers and governmental bodies that increasingly advanced AI systems may facilitate cyberattacks.
The U.K. AI Security Institute reported that during testing, Mythos autonomously executed a 32-step simulated attack on a corporate network. Mozilla indicated that the model identified 271 vulnerabilities in the Firefox web browser which were subsequently fixed, while security startup Calif claimed that a preview version aided researchers in developing an exploit chain targeting Apple’s M5 computer chips.
Anthropic is expanding access to Mythos as it believes that defenders require increasingly advanced AI tools before adversaries can obtain similar capabilities. The company anticipates that other developers may introduce Mythos-class models within the next six to twelve months, potentially lacking equivalent safeguards, which could lead to more frequent and unpredictable cyberattacks.
“In the future, releases of frontier models will carry greater risks. Capabilities will continue to evolve across various fields, including cybersecurity, which can empower both attackers and defenders,” Anthropic remarked. “If we succeed, we aim to provide a lasting advantage for defenders.”
Anthropic did not respond promptly to a request for comment from Decrypt.
With the expansion of private access, users on Myriad—a prediction market platform run by Decrypt's parent company, Dastan—are expressing doubts about the public release of Claude Mythos by the end of June. Currently, they assign just a 26.5% chance as of now, a decline from a high of 58.5% noted last week.