Firefox’s 423 Fixes Expose the Maintainer Gap

In April 2026, Mozilla reported 423 Firefox security fixes, versus 31 a year earlier—even as model-assisted audits were turning up bugs faster than Firefox typically received them.

Turning an automated finding into a fix still requires verification, coordinated disclosure, patch review, and release engineering.

Independent audits outran Firefox’s old reporting cadence

Mozilla said Claude Opus 4.6 found more than 100 Firefox bugs during two weeks in January, including 14 high-severity issues. Mozilla described that as more findings than Firefox typically receives in over two months. OpenAI separately said GPT-6 Cyber found more than 81 novel Firefox vulnerabilities as it prepared a limited launch to Daybreak Red customers. These are separate reported audits, not a combined count of distinct unfixed flaws.

Cloudflare also tested Anthropic’s Mythos Preview against more than 50 of its own repositories and described the harness it used to search for vulnerabilities. Mozilla, OpenAI, and Cloudflare did not need a common research program to make repeated model-assisted inspection worthwhile. They each had code worth protecting and access to tools that could examine it at scale.

OpenAI’s tally has important limits. Its report does not establish that the Firefox flaws were exploited in the wild, that they were equally severe, or how long Mozilla would need to patch them. OpenAI initially restricted GPT-6 Cyber to Daybreak Red customers, so its result does not show that the same capability was broadly available. It does show why a browser maintainer has to be ready to receive findings in batches rather than at its customary pace.

Mozilla raised the rate of shipped fixes, not just reports

Mozilla put 271 fixes for vulnerabilities identified with early access to Mythos Preview into Firefox 150. It then reported that Mythos Preview and other AI models helped it identify and ship 423 Firefox security fixes in April, compared with 31 a year earlier. The 271 fixes belong to that broader April account; adding the two figures would count the release work twice.

Mozilla’s April releases complicate the claim that AI merely moves the bottleneck downstream. Its team characterized the transition as difficult rather than permanently destabilizing. The comparison does not tell us how severe all 423 issues were or establish that every project can reproduce Mozilla’s workflow.

Other builders are treating repair as part of the capability, too. OpenAI and Paradigm designed EVMbench to evaluate agents on detecting, exploiting, and patching smart-contract vulnerabilities. Microsoft introduced Project Perception as an agentic system intended to patch flaws. Neither announcement measures Firefox’s outcomes, but both place the useful unit of security work beyond the initial bug report.

Private disclosure buys time that public knowledge no longer does

Anthropic researchers said Mythos Preview could turn a publicly disclosed vulnerability into a working exploit in hours rather than weeks. That claim concerns known flaws, not proof that attackers used Mythos against Firefox. It nevertheless changes what a disclosure delay is for: maintainers need a protected interval in which to reproduce a report, assess its severity, coordinate with affected parties, and ship a fix.

GitHub made private vulnerability reporting generally available in 2023 so researchers could contact open-source maintainers confidentially. The channel matters more when finding and exploiting a disclosed flaw both get faster. A private inbox alone, however, cannot review a patch or decide when users can safely receive it.

NIST faced a backlog of more than 25,000 vulnerabilities in its National Vulnerability Database in 2025, roughly ten times its previous 2017 high. NIST’s cataloging queue is not Mozilla’s patch queue. Discovery can increase pressure on several different human-operated systems, and clearing one queue does not clear the others.

Smaller dependencies cannot borrow Mozilla’s release team

The XZ Utils backdoor brought attention in 2024 to security risks created by demands on under-supported open-source developers. AI-assisted auditing adds a different burden: a maintainer may receive more plausible reports without gaining more time to validate them. The Linux Foundation said Anthropic, Amazon, Google, Microsoft, and OpenAI supplied $12.5 million in grants to help free and open-source maintainers handle AI-generated security findings. The grants acknowledge that someone must do the work after the scan; they do not establish that every dependency now has sufficient capacity.

Low-quality automated submissions can also consume bug-bounty reviewers’ attention meant for genuine flaws. A small project must distinguish a reproducible vulnerability from a convincing false alarm before it can make a release decision. More model access may help with that task, but access does not itself supply an accountable maintainer.

Browser alliances can share trust decisions, not patch labor

Mozilla, Google, Microsoft, and Cloudflare partnered on Private Access Control Tokens to distinguish legitimate human or bot web traffic from unwanted requests. PACT addresses a different problem from Firefox vulnerability repair. It illustrates how browser security also depends on shared rules for automated activity, while each project remains responsible for the code and releases it controls.

Mozilla’s 423 April fixes show that a browser team can widen its release pipeline when model-assisted discovery accelerates. A small dependency cannot borrow that pipeline merely because a model found its bugs. As long as both defenders and attackers can automate the search, the consequential resource is the maintainer time that turns a private, verified report into a safe release.