Mozilla’s latest security experiment suggests a turning point in how software is secured, with an AI system uncovering problems that years of human effort and conventional tools had missed.

Early scans and rising capabilities

Mozilla has long relied on traditional security testing tools like fuzzers to probe Firefox’s vast codebase, but these methods left some flaws undiscovered for years. When the organization first tried a general-purpose AI model for code review, it surfaced just 22 security‑sensitive bugs in a scan, offering only modest gains over existing techniques.¹

In April 2026, Mozilla shifted to Anthropic’s purpose‑built vulnerability research model, Claude Mythos Preview, and aimed it at Firefox. The company reported that Mythos found “hundreds of issues,” including at least one bug that had persisted for 20 years and gone “undetected for years by fuzzers,” underscoring clear limitations in legacy tooling.¹

April’s surge in fixes

The Mythos‑assisted push led to fixes for 423 security bugs shipped in Firefox’s April releases, with 271 issues specifically attributed to the new AI system — a dramatic jump from the 25 fixes recorded in January and 76 in March.¹ Mozilla described the newer models as far more capable of navigating sprawling browser code and delivering “much more comprehensive coverage” than past AI attempts, which had often produced plausible‑sounding but incorrect recommendations.¹

Rethinking trust and the future role of humans

Commentary on the experiment frames it as more than a cybersecurity milestone. One analysis notes that the previous AI scan found 22 security‑sensitive bugs, while the Mythos run uncovered 271, and argues this is “the first serious sign” that the old assumption — humans write code, machines only check it — “is about to flip.”² In this view, future “serious software” may be generated, attacked, repaired, and verified largely by machines, with humans shifting toward defining what systems are allowed to do and mean.²

That shift, advocates say, could raise the standard for what counts as safe code, making adversarial machine review an expectation rather than an experiment — and turning comprehensible, well‑structured code into a core security property in its own right.²