Anthropic announced on Tuesday the launch of a new initiative in collaboration with multiple technology giants to proactively address potential cybersecurity threats posed by more advanced AI models. The company introduced a cooperative program called "Project Glasswing," through which it will provide an unreleased AI model named "Mythos" to several technology firms, including Amazon.com, Apple, Microsoft, and Cisco. Participating companies will use the model to identify security vulnerabilities in their products and share research findings across the industry. Anthropic stated that there are currently no plans to make Mythos publicly available, and future safety measures for the model will be refined based on feedback from the Glasswing initiative. This move reflects growing concerns within the tech industry that sophisticated AI models could be exploited by hackers or state-sponsored attackers. It is widely acknowledged that as AI capabilities advance, its proficiency in code analysis and vulnerability detection also improves, potentially enabling it to circumvent existing cybersecurity defenses. Anthropic's competitor, OpenAI, has previously highlighted similar risks and launched pilot programs prioritizing the distribution of AI tools to defensive actors. Newton Cheng, head of cybersecurity at Anthropic, emphasized that this is not an issue for any single company but a challenge requiring collaboration across the industry and with government entities. "Through Glasswing, we aim to give defenders a head start before the technology becomes widely accessible," he said. The company also revealed that it has engaged with relevant U.S. government agencies regarding Mythos’ security capabilities and is collaborating with the Cybersecurity and Infrastructure Security Agency (CISA) and the National Institute of Standards and Technology (NIST). Although Mythos was not specifically developed for cybersecurity, its performance has already drawn attention. Anthropic reported that the model has identified several critical vulnerabilities, including a 27-year-old flaw in key internet software and a gaming code defect that went undetected for 16 years, despite being scanned over 5 million times by automated testing tools. To prevent misuse, Anthropic has implemented strict access controls for participants in the Glasswing project, though specific details were not disclosed.
Comments