Anthropic AI assists Calif researchers in identifying macOS M5 silicon vulnerability
Researchers from Palo Alto-based firm Calif, working with Anthropic’s Claude Mythos Preview, claim to have developed a privilege escalation exploit for Apple’s latest hardware, prompting direct engagement with the Cupertino-based tech giant.
Security researchers from the Palo Alto-based company Calif have claimed to have breached macOS by developing a privilege escalation exploit with assistance from Anthropic’s Claude Mythos Preview. The exploit reportedly grants access to inaccessible segments of a MacBook, effectively allowing an attacker to take full control of the computer. The work was conducted as part of Anthropic’s Project Glasswing, an initiative launched in April to prevent AI-driven cyberattacks by leveraging artificial intelligence for defensive security.
Apple has responded to the claims with urgency, meeting with the researchers at its Apple Park campus to discuss the findings. The company described the vulnerability as the first public macOS kernel memory corruption exploit targeting M5 silicon. In a statement to The Wall Street Journal, Apple emphasised that security remains its top priority and that it takes reports of potential vulnerabilities very seriously. The researchers have indicated they will release full technical details only after Apple has patched the identified vulnerabilities.
The collaboration demonstrates how advanced AI systems can accelerate the identification of security flaws. According to reports, Mythos Preview was able to quickly identify the bugs because they belonged to known classes of errors. However, human expertise remained essential for designing the actual exploit, underscoring that while AI can unearth attack paths previously unknown, human oversight is still required to operationalise the findings. This capability raises concerns that such tools could be repurposed by malicious actors to stage security breaches.
Project Glasswing involves a broad coalition of technology and financial institutions, including Amazon Web Services, Microsoft, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, NVIDIA, and Palo Alto Networks. Participants utilise Mythos to strengthen the security of their own projects. Mozilla, for instance, previously announced that it identified and patched 271 vulnerabilities in its latest Firefox browser release with assistance from the same AI model.
The development comes amidst a broader industry shift towards integrating AI into cybersecurity defence. Just days prior to this disclosure, OpenAI introduced its own initiative, Daybreak, which utilises its AI models, including the specialized security agent Codex. Daybreak is built on the premise that cyber defence should be embedded into software development from the outset, rather than relying solely on the detection and remediation of vulnerabilities after the fact.
