AI is getting frighteningly good at finding flaws in software, according to one company that got early access to Anthropic’s powerful new Mythos model.
On Thursday, Mozilla, the nonprofit that develops the Firefox web browser, posted a blog detailing how it used Anthropic’s unreleased AI model, Claude Mythos Preview, to uncover security issues.
The software company said its AI model discovered hundreds of issues, some of which had been around for at least 20 years.
Mozilla detailed the 12 bugs it fixed. One of them “went undetected for years” by Fuzzer, a security testing tool that attempts to break software.
Overall, Mozilla says its AI-assisted efforts helped fix 423 security bugs shipped in the April release of Firefox, including 271 issues that the company specifically tied to the Claude Mythos Preview.
This is a big change compared to previous monthly updates. Mozilla reportedly fixed only 25 bugs in January of this year. According to the company, 76 cases were revised in March.
“The model’s capabilities have been significantly improved,” the company wrote on its blog.
According to Mozilla, previous generations of AI bug reports often resulted in “unwanted slop.” The solution suggested by the AI sounded plausible, but it turned out to be wrong.
Mozilla says the new system can navigate the browser’s sprawling codebase and identify deeply buried flaws. Mozilla said the new generation model allows for “much more comprehensive coverage” than previous AI models.
Mythos functionality has been reported and is prohibited for public use. Anthropic said the model is so powerful that exposing it to the world could pose real national security and online crime risks.
As a result, only a select few companies are mastering the new model. Mozilla is one of them.
The company said it plans to further expand the system by integrating AI analytics directly into the Firefox development pipeline.
Mozilla did not immediately respond to Business Insider’s request for comment.