Firefox is one of the most thoroughly tested browsers in the world. Two decades of security audits, thousands of hours of fuzzing, top researchers from around the globe. And yet, the Mythos Preview model went through the source code and, in a short time, brought to light 271 security vulnerabilities that had been waiting there unnoticed for an attacker or AI. This time, AI got there first.
What actually happened and where it began
It all began in February, when Anthropic's Frontier Red Team started working on Firefox. Their approach was different from traditional penetration testing from the outset. Instead of reviewing the code manually, they deployed Claude, then still the Opus 4.6 model, and had it analyze Firefox's JavaScript engine.
The results came quickly. Anthropic provided Mozilla with test cases that allowed engineers to immediately reproduce and verify each bug. Fixes began within hours. In total, this collaboration uncovered 14 high-severity bugs and led to the issuance of 22 CVEs (security vulnerability identifiers), which were fixed in the March release of Firefox 148.
For comparison, Firefox undergoes some of the most intensive security testing on the web. Yet the model found issues that fuzzing, static analysis, and human researchers had failed to uncover for years.
Mythos found another 271 vulnerabilities
Following the March results, Mozilla continued its close collaboration with Anthropic and gained access to a new, then-unreleased version of the model called Mythos Preview. This time, they deployed it on Firefox version 150. The resulting number made the team dizzy, in their own words. 271 vulnerabilities in a single round of testing on software that has been continuously tested for more than twenty years.
“For a target this hardened, a single such bug would have been a serious signal last year. So many at once makes you stop and wonder whether it is even possible to keep up,” wrote Bobby Holley, Firefox's chief technology officer, in a blog post. Mozilla subsequently fixed all the bugs that were found. Firefox 150, released this week, is therefore the most secure version of the browser in its history.
But Mythos was not built as a security tool
Anthropic's Mythos Preview is not a specialized vulnerability scanner. It is a general-purpose coding model. Its ability to find security flaws is not the result of deliberate design, but a by-product of the model's ability to read and reason about code extremely well.
Mythos can hold entire large codebases in context, trace code execution paths, and identify discrepancies between what the code does and what the developer intended. This exact capability, which until recently was the exclusive domain of elite security researchers, now works automatically.
“Computers could not do this at all a few months ago, and now they excel at it,” says Holley. “We have many years of experience analyzing the work of the world's best security researchers, and Mythos Preview is even more capable.” In other words, what once required months of expensive human labor can now be done by the model in a fraction of the time and at a fraction of the cost.
Why this benefits defenders more than attackers
For years, cybersecurity operated on the principle of a costly balance. An attacker needed to find only one vulnerability. A defender had to protect everything. Companies such as Mozilla therefore tried to make attacks expensive enough that only state actors with virtually unlimited budgets could afford them—and even they would think twice before actually using such an expensive weapon. Mythos complicates this calculation. But not in the attackers' favor.
“The gap between what a machine can discover and what a human can discover favors the attacker,” Holley explains. “By narrowing that gap, we deprive attackers of their long-term advantage because we make every discovery cheaper.” Once defenders and attackers can use the same tools, the winner will be whoever fixes more issues first. And defenders have an incentive to patch; attackers do not. Holley therefore declared in bold text: “Defenders finally have a chance to win.”
At the same time, he emphasized one reassuring point: none of the 271 bugs found was of a kind that an experienced human researcher could not also have discovered. Speculation that AI will uncover entirely new categories of vulnerabilities beyond the grasp of the human mind has not yet been confirmed. “Software like Firefox is designed modularly so that people can reason about its correctness. It is complex, but not arbitrarily complex,” Holley writes.
Open source code is first in line. And it is in the worst position
Firefox is open source. Anyone can download the source code and inspect it. This is excellent for transparency, but it also means that anyone can analyze publicly available code in the same way as Mythos or similar models. Attacker and defender alike.
Mozilla is aware of this and is actively sharing knowledge and tools with other free software projects. Because the problem is broader. The modern internet relies on thousands of libraries and tools maintained by one or a handful of volunteers in their spare time. These people do not have access to Mythos. Nor do they have the resources to process hundreds of bug reports, even if they received them.
“The programmer who has devoted 20 years of their life to maintaining code that runs inside products used by billions of people does not yet have access to Mythos. They should.” Raffi Krikorian, Mozilla's chief technology officer, wrote for the New York Times.
Bobby Holley confirms that the situation will force large companies to change course. “I've spoken with chief technology officers at very large companies who say they are pulling thousands of engineers away from everything else to focus on this for six months,” he says. But small projects and volunteer maintainers simply do not have that kind of capacity.
Access to Mythos is not available to everyone
Anthropic released Mythos Preview only to a closed group of twelve partner companies and expanded access to more than 40 additional organizations that maintain critical software infrastructure. The model is not yet available to the public. Under the umbrella of the Project Glasswing initiative, these include Microsoft, Apple, and Amazon, among others.
The reason Anthropic has not released the model broadly is unusual. In the company's own words, the model is too capable to be safely released to everyone.
Mozilla, however, is not formally part of Project Glasswing. It established a direct collaboration with Anthropic and was among the first to undergo what Holley describes as an inevitable transition for every piece of software. “Every piece of software will have to go through this, because every piece of software has a large number of bugs buried beneath the surface that can now be detected,” he says.
Sources: theregister.com and wired.com
