AI is getting scary good at finding software flaws, according to one company that got early access to Anthropic’s powerful new Mythos model.
On Thursday, Mozilla, the nonprofit behind the Firefox web browser, posted a blog detailing how it used Claude Mythos Preview — an unreleased AI model from Anthropic — to uncover security issues.
The software company said the AI model discovered hundreds of issues, including at least one that had survived for 20 years.
Mozilla detailed 12 bugs it fixed. One, it said, went “undetected for years by fuzzers,” the security-testing tools that try to break software.
Overall, Mozilla said the AI-assisted effort contributed to fixes for 423 security bugs shipped in April releases of Firefox, including 271 issues the company specifically tied to Claude Mythos Preview.
That’s a massive change compared to previous monthly updates. In January this year, Mozilla says it made just 25 bug fixes. Seventy-six were fixed in March, the company said.
“The models got a lot more capable,” the company wrote in the blog.
Earlier generations of AI bug reports often produced “unwanted slop,” Mozilla wrote. The AI-proposed solutions sounded plausible, but turned out to be incorrect.
The newer system, according to Mozilla, could navigate sprawling browser codebases and identify deeply buried flaws. Mozilla said the new generation models allowed for “much more comprehensive coverage” than previous AI models.
Mythos’ reported capabilities have kept it shuttered from public use. Anthropic has said the model is so powerful that releasing it to the world could create real national security and online crime risks.
So, only a handful of chosen companies have gotten use to the new model. Mozilla is one.
The company said it plans to expand the system further by integrating AI analysis directly into Firefox’s development pipeline.
Mozilla did not immediately respond to a request for comment from Business Insider.
Read the full article here
