Researchers have used Anthropic’s Claude Mythos Preview to find security flaws in Apple’s macOS, according to The Wall Street Journal and a blog post from Palo Alto-based security firm Calif.
The case has drawn attention because macOS is widely viewed as a secure operating system. However, the researchers said the AI tool helped identify a complex attack path involving more than one flaw. Apple has received the findings and is now reviewing the report.
Anthropic’s Claude Mythos Preview is an AI security tool built to find software vulnerabilities. The company has not released it to the public. Instead, it has given access to selected security researchers and large organizations.
Calif researchers said Mythos helped them find a macOS kernel memory corruption exploit on Apple M5. The team described it as the ‘first public macOS kernel memory corruption exploit on Apple M5.’ However, full technical details remain private until Apple addresses the reported attack path.
The researchers said the exploit could allow an unprivileged local user to gain broad access to a device. However, the report also made clear that the attack was not a fully independent AI action. Skilled security experts guided the process and handled key parts of the exploit work.
According to the report, Mythos did not rely on one simple coding error. Instead, it helped identify two separate vulnerabilities that could work together in a chained attack.
By linking those flaws, the researchers said the attack could corrupt memory inside macOS. That type of weakness can allow access to protected areas of a system. However, the researchers have not released the full method to avoid giving attackers a ready path before a fix is available.
Calif said the exploit involved ‘two vulnerabilities and several techniques.’ The firm also said Mythos assisted with bug discovery and exploit development. Still, the team stressed that human researchers remained part of the process.
The case has raised doubts about how quickly AI tools can change security research. Mythos found bugs from known bug classes, according to Calif. The firm said, “Mythos Preview is powerful: once it has learned how to attack a class of problems, it generalizes to nearly any problem in that class.”
That statement points to a key concern in cybersecurity. AI systems may help researchers review large codebases faster than before. However, the same speed could also help attackers if similar tools reach the wrong users.
Reports also noted that Mythos did not complete the full attack alone. Security professionals still needed to guide the final steps. Therefore, the case shows both the strength and limits of AI-assisted hacking.
Apple has not confirmed whether the specific flaws found by Calif have already been fixed. Reports noted that macOS Tahoe 26.5 release notes mention a bug submitted by Calif in collaboration with Claude and Anthropic Research. Calif also appeared in other vulnerability reports.
However, Calif said it met with Apple early in the week and would share full technical details after Apple fixes the vulnerabilities and attack path. That timeline has created doubts over whether all related issues have already been patched.
An Apple spokesperson told The Wall Street Journal, “Security is our top priority, and we take reports of potential vulnerabilities very seriously.”
For now, Apple is reviewing the data provided by the research team. The limited public details mean users cannot see the full attack chain. However, the report shows that AI-assisted vulnerability research is moving into more secure software environments, including macOS.
Also Read: White House to Provide Anthropic Mythos Access to US Agencies Despite Pentagon Rift