Security Researchers Claim AI-Assisted Exploit Targets Apple Mac M5 Kernel via Anthropic Claude Preview
A security research team affiliated with startup Calif has publicly claimed it successfully developed an Apple macOS kernel exploit targeting the Mac M5, with the attack chain reportedly aided by Anthropic's preview release of its Claude Mythos AI model. The researchers outlined their methodology at a security conference, asserting that large language model technology lowered barriers to constructing sophisticated kernel-level attacks against Apple hardware. The disclosure has renewed scrutiny over how frontier AI capabilities intersect with offensive security research and whether preview models carry adequate safeguards against misuse.
The specific details of the exploit remain under close examination by the broader security community. According to the presenters, the Claude Mythos preview assisted in identifying code paths and constructing payloads targeting macOS kernel interfaces, a realm traditionally requiring deep expertise in low-level system programming. While the security field has long debated whether AI could meaningfully accelerate exploit development, this claim represents one of the more concrete demonstrations of LLM-assisted vulnerability research targeting a major consumer platform. Apple has not publicly responded to the specific allegations as of this reporting.
The incident raises pressing questions for both AI developers and platform vendors. Anthropic and other frontier labs maintain usage policies intended to discourage harmful applications, yet researchers argue the boundary between legitimate security research and misuse remains contested terrain. For Apple, the disclosure adds pressure to its already high-stakes security posture, particularly as the Mac M5 represents the company's latest generation of custom silicon. The research underscores an accelerating arms race between defensive teams and adversaries leveraging AI to compress timelines for vulnerability discovery and exploitation.