Tech • AI • Robotics
He Gained Access to Mythos, the AI Forbidden to the Public
9.4/10TL;DR
Mozilla engineers used Anthropic’s Mythos model to hunt vulnerabilities in Firefox, uncovering hundreds of serious flaws and intensifying concerns that AI is dramatically accelerating both cyberdefense and cyberattacks.
KEY POINTS
Browsers remain a prime target
Browsers are among the most attractive targets in cybersecurity because they serve as the main gateway to computers and phones. They process untrusted HTML, CSS, JavaScript, WebAssembly, audio, video and network content in real time, creating a vast attack surface. Threat actors range from criminal groups to intelligence services because a browser exploit can potentially expose data or open a path to full device compromise.
A long history of severe exploits
One previously reported browser flaw allowed a malicious news site viewed in Russia to extract locally stored password files and send them to a server in Ukraine during the Crimea period. Such examples show how a simple click can lead to credential theft or remote control of a machine. In browser security, the most prized exploits are often zero-click or one-click attacks that require little or no user interaction.
Securing a browser is unusually difficult
Firefox contains roughly 33 million to 34 million lines of code, much of it written in C and C++, languages long associated with memory-safety problems. Roughly 50% to 60% of security flaws across major software projects stem from memory management errors. Browsers also change constantly, with hundreds of code changes per day and new features such as WebRTC expanding the surface for regressions and vulnerabilities.
Traditional defenses were already stretched
Before the latest AI push, security teams relied on static analysis, bug bounty programs and especially fuzzing, which bombards software components with valid and malformed inputs to trigger unexpected behavior. Fuzzing has become one of the most effective automated methods for finding vulnerabilities, producing hundreds of findings per year. But it does not catch every class of bug, especially complex logic flaws or deeper exploit paths.
Mythos quickly proved more than marketing
Early contact with Mythos reportedly began after Anthropic sent high-quality vulnerability reports affecting Firefox. That immediately signaled that the model was finding meaningful issues rather than low-value noise. Engineers were then granted direct access so they could run the model against their own codebase with tailored attack prompts and verification steps.
Hundreds of serious flaws surfaced
Security teams said the model uncovered hundreds, and close to 1,000, vulnerabilities across the browser. Instead of fixing around 20 new vulnerabilities in a month, engineers found themselves correcting more flaws than they would normally address in a year. These were not cosmetic bugs: they were classified as real security issues that could potentially allow compromise of the browser, even if not all led to sandbox escape.
Some bugs had been hidden for years
Engineers found vulnerabilities that had apparently existed for 15 to 20 years in legacy code. The scale of the discovery far exceeded prior estimates, with one internal estimate of roughly 50 to 60 important unknown flaws suddenly looking off by an order of magnitude. Similar patterns were reportedly visible beyond Mozilla, suggesting the issue is industry-wide rather than specific to one browser.
Expertise still matters
Results depended heavily on how the model was used. General prompts such as asking for a generic memory bug were far less effective than specialized prompts written by engineers with decades of browser security experience. That helps explain why some public tests on other projects appeared underwhelming at first, only for more informed researchers to later find many additional flaws.
A temporary edge for defenders
Mozilla integrated AI-assisted vulnerability hunting into the same continuous pipelines used for fuzzing, letting new code be tested within hours. The effort involved about 100 engineers for several weeks and appears to have paid off: at Pwn2Own Berlin, where successful browser exploits can earn between $50,000 and $150,000, several expected attacks against Firefox failed after a large wave of fixes landed.
The next risk is patch-speed asymmetry
The larger concern is no longer just finding bugs, but how quickly attackers can weaponize disclosed fixes. AI models can already inspect code changes and infer whether a patch addresses a security flaw, then help generate an exploit before all users update. Automatic updates and rules such as the EU Cyber Resilience Act improve the situation, but many large organizations still cannot roll out patches in hours, leaving a dangerous window for exploitation.
CONCLUSION
AI-assisted security research is rapidly exposing a backlog of serious software flaws, offering defenders a chance to harden critical systems faster than before. But unless patching and internal containment improve just as quickly, the same acceleration could shift the advantage back to attackers.
Explain this