Mozilla reports that Anthropic’s Mythos identified 271 previously unknown vulnerabilities in the Firefox browser, far exceeding the results of earlier AI model testing.

Since February, Mozilla has conducted extensive testing of its Firefox browser using frontier AI models to detect cyber vulnerabilities. In its latest round of testing, Mozilla discovered 271 previously unknown vulnerabilities using Anthropic’s “Mythos.”

On April 7, Anthropic limited the release of Claude Mythos2 Preview, or “Mythos,” to a select list of industry partners for cybersecurity testing. Anthropic claimed that Mythos worked so well at identifying cyber vulnerabilities that releasing it to the public was dangerous. Under the umbrella of Anthropic’s Project Glasswing, Mozilla was granted access to Mythos for cybersecurity testing.

Mozilla previously did similar testing on an earlier browser version, Firefox 148, using Claude 4.6, the previous AI model from Anthropic. Claude 4.6 identified cybersecurity risks, leading to 22 fixes – significantly less than Mythos’ findings. Mozilla said all identified issues are being reviewed, and confirmed vulnerabilities are being addressed through its existing security process.

The company said the test is part of its broader effort to evaluate how AI systems can be used to identify software vulnerabilities.