Mozilla Turns AI on Its Own Browser: 271 Latent Bugs Fixed in Firefox 150

IPC races appeared repeatedly. One allowed a use-after-free in IndexedDB refcounting that could lead to sandbox escape. Another passed a raw NaN value that masqueraded as a JS object pointer across process boundaries. WebTransport flooding triggered parent-process use-after-free. DNS parsing produced buffer over-reads when fed crafted HTTPS RR and ECH records via a simulated server.

Even the color picker wasn’t safe. A use-after-free surfaced through nested event loops and garbage collection. HTML table rowspan attributes accepted values over 65,535, bypassing clamps and causing overflows. RLBox sandboxing had a verification gap in value copying.

These weren’t theoretical. Many carried sec-high ratings. Of the 271 AI-discovered issues, 180 rated sec-high, 80 sec-moderate, and 11 sec-low. Mozilla credited three CVEs to Anthropic: CVE-2026-6746, CVE-2026-6757, and CVE-2026-6758. The full set shipped in Firefox 150, plus point releases 149.0.2, 150.0.1, and 150.0.2.

Yet no single bug hands over the system. Sandboxing, OS address space layout randomization, and other defenses remain. Many exploits require first compromising a content process. The AI attempts often failed exactly because prior hardening blocked them. Prototype freezing, introduced in Bug 1771084, stopped several prototype-pollution escapes cold.

Over 100 people contributed. They wrote code, triaged reports, tested fixes, and pushed releases. The pace was intense. Long days became normal. The payoff arrived in one of the most secure Firefox versions ever released.

TechCrunch covered the implications the same day. Its article noted that Mythos unearthed high-severity bugs dormant for more than a decade and highlighted the jump from 31 fixes the prior April to 423 this year. The TechCrunch report quoted the Mozilla team’s pride in the collective effort.

Earlier coverage from Ars Technica in April had already signaled the scale. When Mozilla first announced the collaboration, Firefox CTO Bobby Holley said defenders now “have a chance to win, decisively.” The Ars Technica story compared the 271 bugs to the 22 found previously with Opus 4.6 on Firefox 148.

The Mozilla Blog post from April 21 framed the work as shifting the balance against zero-days. It described round-the-clock efforts since February using frontier models. Fixes landed before public release. Risk of exploitation stayed minimal. That Mozilla Blog announcement emphasized preparation for even stronger future models.

Security advisories confirm the breadth. The Mozilla Foundation Security Advisory for Firefox 150 lists numerous memory safety bugs presumed exploitable for arbitrary code execution with sufficient effort. Similar notices cover the point releases and ESR branches.

Industry observers reacted quickly on X. Discussions highlighted the near absence of false positives in the pipeline and the jump in monthly fixes. One thread linked back to the Hacker News discussion that surfaced the same day the Mozilla Hacks article dropped.

The approach offers lessons beyond browsers. Mozilla plans to fold analysis into continuous integration so every patch gets scanned. The harness already deduplicates against known issues. It articulates pathology clearly enough for engineers to act fast. And it scales with better models.

Anyone building software can start today, the authors advise. Modern models combined with a test harness deliver immediate value. Waiting for perfection isn’t necessary. The current moment carries risk but also opportunity. Firefox now demonstrates what focused application of these tools can achieve.

Memory safety bugs still dominate the list. Use-after-free, boundary errors, integer overflows. Yet the pipeline caught logic errors too. Race conditions in event loops. Hash table reentrancy in XSLT after 20 years. Wallpaper image decoding that bypassed sandbox expectations in the parent process.

Such findings reinforce a basic truth. Complex code accumulates blind spots. Human review misses them. Fuzzers catch some. Static analysis flags others. But an agent that writes tests, runs them, and explains the fault finds patterns humans overlook.

Firefox 150 isn’t perfect. No software is. The team acknowledges that. They continue internal discovery alongside external reports and traditional fuzzing. The 271 represent one slice of a larger April total that also included 41 external bugs and 111 internal ones from other sources.

Still, the acceleration stands out. From dozens of fixes per month to hundreds. From reactive patching to proactive discovery at scale. The combination of agentic workflows, reproducible testing, and frontier models changes the economics of hardening large codebases.

Enterprise users and security teams should take note. Update promptly. Firefox 150 and the associated ESR releases close doors that had remained open far longer than expected. The patches address real attack primitives even if full compromise requires chaining multiple flaws.

For the broader industry, the message is direct. Build the harness. Integrate the models. Start finding bugs before adversaries do. The tools exist now. The data from Mozilla shows they work.