Anthropic says Mythos has already found more than 10,000 vulnerabilities

Trending 1 week ago

The institution has published an update astir Project Glasswing, a period aft its launch.

A man speaking.

Michael M. Santiago/Getty Images

Anthropic has published an first study for Project Glasswing, nan cybersecurity inaugural it launched successful April that intends to prevent AI cyberattacks with, well, AI. The inaugural is powered by Claude Mythos Preview, nan company's unreleased model, which Anthropic says has already helped its partners find more than 10 1000 vulnerabilities wide conscionable a period aft Glasswing's launch. In addition, it says astir of its partners person "each recovered hundreds of critical- aliases high-severity vulnerabilities successful their software" utilizing nan model. 

The institution said that its partners' complaint of bug-finding has accrued by much than a facet of ten. Cloudflare recovered 2,000 bugs, 400 of which are precocious aliases captious successful severity. Mozilla antecedently reported that it recovered and fixed 271 vulnerabilities successful Firefox, 10 times much what it recovered successful an older type of nan browser utilizing different Claude model. 

Microsoft's recent announcement that its spot releases will "continue trending larger for immoderate time" is apparently because of nan bugs it recovered done Mythos Preview. Anthropic besides utilized Mythos Preview to scan 1,000 open-source projects complete nan past fewer months and recovered 6,202 high- and critical-severity vulnerabilities retired of 23,019. While nan institution didn't see it successful nan report, a information investigation patient precocious claimed that it found a measurement to breach macOS, an operating strategy known for having tight security, pinch thief from Mythos' bug-finding capabilities. 

The institution explained successful its study that it hasn't released Mythos Preview to nan nationalist yet, because nary institution (including itself) has developed safeguards beardown capable to forestall models for illustration it from being misused. It intends to merchandise "Mythos-class models" successful nan future, though, erstwhile those safeguards go available. For now, it's readying to activity pinch partners for illustration nan US and different governments to grow nan readiness of Project Glasswing. That indicates that nan institution whitethorn beryllium connected its measurement to repairing its narration with nan US government. The institution is already moving pinch respective partners astatine nan moment, including Amazon Web Services, Apple, CrowdStrike, Google, JPMorganChase, NVIDIA and Palo Alto Networks, successful summation to nan others we've already mentioned. 

Anthropic is reportedly astir to beryllium profitable for nan first clip since it was founded successful 2021. According to a recent report by the The Wall Street Journal, it's on way to station a gross of $10.9 cardinal pinch an operating profit of $559 cardinal for nan 4th ending successful June. The institution doesn't expect to stay profitable successful nan quarters that will follow, however, arsenic it intends to put much money into computing resources and different expenses. 

More
Source engadget.com
engadget.com