Mozilla researchers said Anthropic’s Mythos AI model has uncovered numerous high-severity vulnerabilities in the Firefox browser, including flaws that had remained undetected in the software for more than a decade.
The findings provide one of the clearest public examples so far of how advanced AI systems are changing software security and vulnerability research.
Anthropic introduced the Mythos model in April and said at the time that the system was capable of identifying software vulnerabilities at a level significant enough that thousands of severe bugs would need to be addressed before the model could be released publicly.
In a post published Thursday, Mozilla researchers described how Mythos substantially improved the company’s internal bug-finding efforts compared with earlier AI security tools.
Previous generations of AI vulnerability scanners often overwhelmed security teams with false positives and low-quality reports. Mozilla said the newer systems have become far more reliable because agentic AI tools can now evaluate and filter their own findings more effectively.
“It is difficult to overstate how much this dynamic changed for us over a few short months,” Mozilla researchers wrote.
“First, the models got a lot more capable. Second, we dramatically improved our techniques for harnessing these models,” the researchers added.
Firefox Bug Fixes Increased Sharply
Mozilla said the impact of the AI-assisted security work has already become visible in Firefox development.
The browser shipped 423 bug fixes in April 2026, compared with 31 fixes during the same month a year earlier.
Researchers also published technical details on 12 vulnerabilities identified through the process. The issues included two sandbox-related vulnerabilities and a 15-year-old flaw involving how Firefox parsed a specific HTML element.
Brian Grinstead said the current generation of AI systems has improved significantly compared with tools available only months earlier.
“These things are actually just suddenly very good,” Grinstead told TechCrunch.
He said Mozilla has observed the improvements across internal scanning systems, external bug reports, and wider industry activity.
Sandbox Vulnerabilities Highlight AI Capabilities
Mozilla researchers said the AI system’s ability to identify Firefox sandbox vulnerabilities was especially notable because of the complexity involved in detecting such issues.
Sandbox exploits often require multiple coordinated steps, including creating compromised browser code and then attacking highly secured portions of the software environment.
Mozilla’s bug bounty program offers rewards of up to $20,000 for researchers who identify Firefox sandbox vulnerabilities, representing the highest payout tier in the company’s program.
Despite those incentives, Grinstead said Mythos has uncovered sandbox-related issues at a higher rate than human researchers historically achieved.
“We do get them,” Grinstead said, “but not at the volume that we are able to find with this technique.”
Humans Still Handle Bug Fixes
Although AI systems are increasingly capable of identifying vulnerabilities, Mozilla said engineers are still responsible for writing and reviewing production-ready fixes.
The company said AI-generated code patches currently serve primarily as references or starting points rather than deployable solutions.
“For the bugs we’re talking about in this post, every single one is one engineer writing a patch and one engineer reviewing it,” Grinstead said. “We have not found it to be automatable.”
The broader security implications of increasingly capable AI systems remain uncertain.
Mozilla noted that many vulnerabilities identified by Mythos may still remain unpatched shortly after discovery, while similar techniques could potentially be used by attackers as well as defenders.
Dario Amodei recently said he believes stronger AI vulnerability discovery tools could ultimately improve security by helping companies identify and fix more flaws.
“If we handle this right, we could be in a better position than we started, because we fixed all these bugs,” Amodei said during a recent event.
Grinstead offered a more cautious assessment, saying AI tools currently provide advantages to both defenders and attackers.
“It’s useful for both attackers and defenders, but having the tool available shifts the advantage a little bit to defense,” he said. “Realistically, nobody knows the answer to this yet.”
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.