Security startup Calif claims to have used a preview version of Claude Mythos to build a working exploit against Apple M5’s Memory Integrity Enforcement protections.
Vietnamese startup Calif claims to have developed the first public macOS kernel memory corruption exploit capable of surviving the new Memory Integrity Enforcement (MIE) protections on Apple M5 hardware. The announcement was published in a Substack post on Thursday, May 14, 2026, and states that a small team of researchers completed the work in less than a week, with assistance from a preview version of Anthropic’s Claude Mythos.
According to Calif, the attack path was discovered accidentally on April 25, when researchers identified the bugs, before developing a working exploit by May 1. The exploit chain targets macOS 26 running on Apple M5 systems, starting from an unprivileged local user account and escalating to root access via standard system calls. The exploit combines two vulnerabilities and additional techniques targeting M5 bare-metal hardware with the MIE kernel enabled. Calif shared its findings with Apple in a meeting at the company’s California headquarters.
Claude Mythos Preview helped identify the vulnerabilities and provided assistance throughout exploit development, though Calif emphasizes that human expertise remained necessary to bypass MIE protections. “Part of our motivation was to test what is possible when the best models are paired with experts,” the company wrote. “Building a kernel memory corruption exploit against best-in-class protections in a week is noteworthy, and it says something loud about this pairing.”
Anthropic released the Mythos preview version in April, following internal testing and external evaluations indicating that the model is capable of autonomously identifying and exploiting software vulnerabilities at a level beyond previous public AI models. Rather than making it publicly available, Anthropic restricted access to select technology companies, banks, and researchers as part of its Project Glasswing initiative. That same month, it emerged that the U.S. National Security Agency was using Mythos. Mozilla subsequently reported that Mythos identified 271 vulnerabilities in Firefox during internal testing, while the UK’s AI Security Institute found that the model is capable of autonomously completing sophisticated multi-stage cyberattack simulations.
On the prediction market platform Myriad, users assign only a 10.5% probability to a public release of Claude Mythos by June 30, 2026. Apple has not yet issued any public comment on Calif’s claims. The startup described the Apple M5 exploit as “a taste of what is coming,” adding: “Apple built MIE in a world before Mythos Preview. We are about to find out how the best mitigation technology on Earth holds up during the first AI bugmageddon.”
