Security experts from Calif, a cybersecurity firm based in Palo Alto, have leveraged methodologies from an early form of Anthropic’s clandestine Mythos AI model to identify two previously unknown vulnerabilities in Apple’s macOS.
These vulnerabilities have been combined to create a privilege escalation exploit capable of circumventing Apple’s advanced memory integrity protections, allowing unauthorized access to typically restricted system areas.
Apple is currently assessing Calif’s comprehensive 55-page report, and patches are anticipated following the validation of the findings.
Unveiling macOS Security Gaps
The exploit, identified during testing sessions in April, combines two macOS bugs with several sophisticated techniques to compromise the Mac’s memory, ultimately infiltrating secure system zones inaccessible to ordinary processes.
Reports from The Wall Street Journal indicate that, if combined with other attacks, this exploit could enable a malicious entity to gain complete control over the targeted Mac system.
Calif’s team developed bespoke software to link these vulnerabilities, crafting an attack vector unprecedented in macOS history.
The Role of Mythos AI
Notably, this exploit is not a remotely deployable worm; significant human expertise is required beyond what Mythos AI alone can achieve. Calif CEO Thai Dong emphasized that the attack’s success heavily depended on the cybersecurity acumen of Calif’s team.
Anthropic’s Mythos, initially known as the Claude Mythos Preview, remains unreleased to the public due to its remarkable yet potentially hazardous capability to pinpoint software vulnerabilities.
This model is integral to Anthropic’s broader Project Glasswing initiative, offering controlled access to about 40 selected organizations, including tech giants like Apple, Google, and Microsoft, for defensive security research purposes.
Implications for the Future
Anthropic has pledged up to $100 million in usage credits to foster this collaborative security effort. Prior to identifying the macOS flaws, Mythos had already demonstrated its prowess by uncovering a bug in OpenBSD, undetected for 27 years, and identifying vulnerabilities in Linux systems that posed hijacking risks.
Anthropic engineers have explicitly cautioned that the model’s ability to expose security flaws is too powerful to be released without stringent controls.
Confident in their findings, Calif researchers personally delivered the technical report to Apple’s headquarters in Cupertino.
An Apple spokesperson, in response to The Wall Street Journal, remarked that security remains a top priority and that potential vulnerability reports receive serious consideration.
Although Apple has not confirmed whether it has started addressing the reported issues, Calif CEO Thai Dong expressed optimism that the vulnerabilities will be swiftly resolved. Detailed technical information of Calif’s discoveries will remain confidential until Apple has rectified the core issues.
