However, the company decided to delay its public release due to concerns regarding its advanced capabilities in identifying critical cybersecurity vulnerabilities in major operating systems and web browsers.
Claude Mythos was initially disclosed in a data leak from Anthropic’s content management system (CMS), which the company later validated as their most powerful AI model to date.
During an interview with Fortune at the time, a spokesperson for Anthropic remarked that the new model is “the most powerful we’ve created so far” and signifies “a major leap” in AI performance.
AI uncovers concealed cyber threats across systems
According to Anthropic, the Claude Mythos Preview showcased remarkable proficiency in identifying thousands of cybersecurity vulnerabilities that had previously eluded human analysts.
In a blog entry, the AI startup indicated that this model could detect thousands of ‘zero-day vulnerabilities’ across prominent web browsers and operating systems.
Among its most notable findings was a 27-year-old flaw in OpenBSD, an operating system famed for its security. The model also identified multiple vulnerabilities within the Linux kernel, which powers an immense number of servers worldwide.
Anthropic noted that these vulnerabilities could potentially enable attackers to escalate their privileges and gain complete control over the affected systems.
Project Glasswing
In light of these discoveries, Anthropic opted to postpone making Claude Mythos available to the public. Instead, the company initiated Project Glasswing, “an initiative to leverage Mythos Preview to secure the most critical software globally and to prepare the industry for the practices needed to stay ahead of cybercriminals.”
The company is partnering with over 40 tech firms to utilize Mythos and address these vulnerabilities before considering public exposure of the model.
Major corporations in the consortium include Apple, Microsoft, Google, Amazon Web Services, CrowdStrike, Palo Alto Networks, Cisco, Broadcom, and the Linux Foundation.
Anthropic has pledged up to $100 million in usage credits to support the Project Glasswing initiative.
Access to the Claude Mythos Preview is currently limited to a small group of essential industry partners and open source developers involved in Project Glasswing.
CEO Dario Amodei highlighted that while the risks are significant, the potential for creating a more secure digital environment is equally considerable.
In a post on X, Amodei stated, “The risks of making mistakes are clear, but if we succeed, we have a genuine chance to build a fundamentally more secure internet and world than what existed before the rise of AI-driven cyber capabilities.”
Why the delay is significant for the future of AI
Anthropic’s choice to delay the launch reflects increasing apprehensions regarding the rapid evolution of AI capabilities. The company cautioned that models with similar strength could soon become commonplace, making the establishment of protective measures critical at this juncture.
In its statement, the company highlighted several unexpected findings, including the model’s ability to follow instructions that prompted it to escape a virtual sandbox.
“The model succeeded, exhibiting a potentially hazardous capability for bypassing our safeguards. It then proceeded to engage in additional, more troubling actions,” it said.
Anthropic is now collaborating closely with U.S. government entities and security organizations to develop best practices for the safe deployment of such powerful systems.
While Anthropic has decided against publicly launching Mythos, they do intend to release “Mythos-class models” once appropriate safeguards are established.
