Anthropic Unveils New AI Model to Enhance Cybersecurity
Anthropic is rolling out a new AI model as part of a cybersecurity initiative in collaboration with major tech players like Nvidia, Google, Amazon Web Services, Apple, and Microsoft. This initiative, known as Project Glasswing, aims to help large organizations—and potentially government entities—identify vulnerabilities in their systems with minimal human oversight.
The company is providing its launch partners with access to Claude Mythos Preview, a versatile model that won’t be publicly released due to security concerns. Newton Cheng, the cyber lead for Anthropic’s frontier red team, mentioned that the model is designed to give cyber defenders a “head start” against threats. Partners will utilize this model to analyze their systems for critical vulnerabilities and assist in patching them. Access is limited to prevent adversaries from exploiting it to discover weaknesses and launch attacks.
While Claude Mythos Preview wasn’t specifically developed for cybersecurity, Anthropic claims that its “strong agentic coding and reasoning skills” contribute to its effectiveness in this area. Cheng refrained from disclosing specific details about the model’s cybersecurity achievements beyond publicly available examples, but the company noted that in recent weeks, Mythos Preview has identified “thousands of high-severity vulnerabilities, including some across all major operating systems and web browsers.” Notably, the blog post emphasizes that the model autonomously identified vulnerabilities and developed related exploits without any human intervention.
The existence of Claude Mythos Preview was initially revealed last month due to a data leak, which Anthropic attributes to human error. Dianne Penn, head of product management at Anthropic, stated that the company is “taking steps in terms of solidifying our processes” and clarified that the leak was unrelated to software vulnerabilities.
Mythos Preview will be available privately to the company’s Glasswing partners, which include JPMorgan Chase, Broadcom, Cisco, CrowdStrike, the Linux Foundation, and Palo Alto Networks, among around 40 other organizations involved in software infrastructure. Anthropic plans to subsidize the usage costs, committing up to $100 million in usage credits and $4 million in direct donations to the Linux Foundation and the Apache Software Foundation, according to Cheng. In the long run, as Anthropic and other AI firms face pressure to generate profit, this program could transition into a paid service if it proves effective enough for continued use by companies.
Despite its recent public disagreements with the Trump administration, Anthropic has indicated that it is in “ongoing discussions with US government officials about Claude Mythos Preview and its offensive and defensive cyber capabilities.” When asked for clarification, Penn confirmed that the company had briefed senior US government officials about Mythos and its capabilities, and that they remain “committed to working closely with all different levels of government.” Cheng noted that while Anthropic is “engaged with” the government, he did not specify who had been briefed.
