Anthropic limits Mythos AI rollout over fears hackers could use model for cyberattacks

Anthropic has officially launched an advanced AI model, named Claude Mythos Preview, as part of its new cybersecurity initiative dubbed Project Glasswing. This model is designed to excel at identifying vulnerabilities and security flaws in software. However, in light of the potential for misuse, the company is restricting access to a select group of companies. Among the initial partners participating in this project are tech giants such as Apple, Google, Microsoft, Nvidia, and Amazon Web Services, all of which will utilize the model for defensive security operations. Additionally, over 40 other companies, including CrowdStrike and Palo Alto Networks, have joined the effort. Dianne Penn, Anthropic's head of research product management, emphasized the company's cautious approach, stating, "There was a lot of internal deliberation. We really do view this as a first step for giving a lot of cyber defenders a head start on a topic that will be increasingly important." The move comes in the wake of reports highlighting the model's advanced cyber capabilities, which sparked concerns in the market and led to a slight drop in cybersecurity stocks. The iShares Cybersecurity ETF experienced minimal fluctuation during trading on the day of the announcement. Founded in 2021 by former OpenAI researchers and executives, Anthropic has worked diligently to build a reputation focused on responsible AI deployment. The announcement of Project Glasswing follows recent public scrutiny over the company’s safety measures following a clash with the Defense Department regarding AI safety practices. Anthropic has been engaged in ongoing discussions with U.S. government officials about the cyber capabilities of Claude Mythos Preview. The project’s name, inspired by the glasswing butterfly, reflects the concept of software vulnerabilities being 'invisible' yet critical. The model has demonstrated its ability to detect previously elusive bugs, including a notable 27-year-old vulnerability in OpenBSD, an operating system renowned for its security emphasis. While Claude Mythos Preview is a general-purpose model not specifically tailored for cybersecurity, its robust coding and reasoning capabilities enhance its effectiveness in identifying security flaws. While Anthropic does not plan to make the model widely available, it aims to explore how it can deploy Mythos-class models on a larger scale in the future. The companies involved in Project Glasswing are tasked with using the model to enhance the security of both proprietary and open-source systems. Anthropic has allocated up to $100 million in usage credits to support these initiatives, although partners will incur costs beyond this limit. Newton Cheng, Anthropic's Frontier Red Team cyber lead, highlighted the importance of allowing these firms to adapt to the capabilities of the model before it becomes broadly accessible. He noted, "Cybersecurity is just going to be an area where this broad increase in capabilities has potential for risk, and thus we have to keep a really close eye on what's going on there."