Anthropic’s Claude Mythos AI Finds 271 Vulnerabilities in Firefox—Yes, It’s Seriously Powerful

For years, opponents have actually had the benefit in cybersecurity. Expert system might will alter that.

In an article released on Tuesday, Firefox internet browser designer Mozilla stated an early variation of Anthropic’s Claude Mythos AI– which has actually drawn attention in current weeks for its supposed cybersecurity expertise– design assisted recognize 271 vulnerabilities in the internet browser throughout internal screening. Those bugs were covered today.

The outcomes highlight how innovative AI systems can evaluate big codebases and find weak points that formerly needed comprehensive manual evaluation by human cybersecurity scientists.

” As these abilities reach the hands of more protectors, numerous other groups are now experiencing the very same vertigo we did when the findings initially entered into focus,” Mozilla composed. “For a solidified target, simply one such bug would have been red-alert in 2025, therefore numerous simultaneously makes you stop to question whether it’s even possible to maintain.”

Mozilla had actually previously evaluated another Anthropic design that determined 22 security-sensitive bugs in a previous Firefox release. In spite of these successes, Mozilla acknowledged that the cybersecurity market has actually long dealt with the total removal of software application exploits as an “castle in the air.”

” Previously, the market has actually mainly battled security to a draw,” the business composed. “Suppliers of vital internet-exposed software application like Firefox take security exceptionally seriously and have groups of individuals who rise every early morning considering how to keep users safe.”

Mozilla stated the brand-new AI system can evaluate source code and recognize vulnerabilities in manner ins which formerly depended upon limited human knowledge. Nevertheless, Mozilla stated the business was motivated to see that no bugs were discovered that could not have actually been found by “an elite human scientist.”

” Some analysts forecast that future AI designs will uncover totally brand-new kinds of vulnerabilities that defy our present understanding, however we do not believe so,” they stated. “Software application like Firefox is developed in a modular method for human beings to be able to factor about its accuracy. It is intricate, however not arbitrarily intricate.”

The outcomes, nevertheless, recommend AI tools might enable designers to reveal great deals of vulnerabilities before opponents exploit them– though on the other hand, in the incorrect hands, it might spell huge problem for software application companies and users alike.

Released in March, Mythos is Anthropic’s most innovative design for thinking, coding, and cybersecurity jobs. Internal business products explain the system as part of a brand-new design tier beyond the business’s earlier Opus series.

Checking carried out before the design’s release revealed it might recognize countless formerly unidentified vulnerabilities throughout significant os and web internet browsers.

Anthropic has actually restricted access to the system through a limited program called Task Glasswing, which provides choose innovation business– consisting of Amazon, Apple, and Microsoft– the capability to utilize the design to scan software application for weak points. It shows a growing effort within the cybersecurity market to utilize AI systems to recognize and spot vulnerabilities before opponents can exploit them.

Nevertheless, the very same innovation might likewise make it possible for brand-new kinds of cyberattacks. Security scientists state AI systems efficient in examining code at scale might automate the discovery of exploitable vulnerabilities throughout commonly utilized software application.

After the launch of Mythos, screening by the U.K.’s AI Security Institute discovered that the AI might autonomously perform intricate cyber operations, consisting of finishing a multi-stage business network attack simulation without human help. Those abilities have actually drawn attention from federal governments and intelligence companies alike.

In spite of a call from President Donald Trump’s administration to stop utilizing Anthropic’s innovation due to a clash over its usage in war and security matters, on Monday, the National Security Company was exposed to be running Claude Mythos Sneak peek on classified networks, according to sources acquainted with the implementation. Using Mythos highlights the growing interest amongst U.S. security companies in the design’s capability to recognize vital software application vulnerabilities.

The design’s efficiency has actually likewise exposed limitations in existing AI examination systems. Previously this month, Anthropic acknowledged that numerous cybersecurity standards are no longer adequate to determine the abilities of its most recent designs.

Mozilla stated the outcomes indicate a possible shift in cybersecurity, where protectors might start to close the enduring benefit opponents have actually held.

” We are exceptionally pleased with how our group increased to fulfill this obstacle, and others will too,” Mozilla composed. “Our work isn’t ended up, however we have actually turned the corner and can look a future far better than simply maintaining. Protectors lastly have an opportunity to win, decisively.”

Mozilla did not instantly react to an ask for remark by Decrypt.