The AI security story finally got out of the slide deck and into a browser release.
Mozilla published a detailed write-up on how it used Claude Mythos Preview and other models to harden Firefox. TechCrunch covered the headline numbers: Firefox shipped 423 security bug fixes in April 2026, compared with 31 in the same month a year earlier.
That is not a normal change in volume.
That is the software security workflow bending.
The old problem
AI bug reports used to be mostly exhausting.
They looked plausible. They wasted maintainer time. They produced noisy claims that humans had to disprove. The cost asymmetry was brutal: cheap to generate, expensive to triage.
Mozilla says that dynamic changed fast.
The models got stronger, but that is only half the story. Mozilla also built a harness around them: prompts, test generation, reproducible cases, deduplication, triage, bug tracking, review, release management, and actual engineers writing fixes.
That last part matters.
This is not “AI found bugs, AI fixed Firefox, everyone went home.”
It is “AI became useful once it was embedded inside a real security pipeline.”
The pipeline is the product
The Mythos work found the kinds of bugs that are hard for normal tooling: multi-step sandbox issues, old parser bugs, race conditions, weird edge cases across distant browser systems.
Mozilla’s post includes examples that had survived years of fuzzing and human review.
But the most important detail is not that the model was clever.
It is that the model could produce reproducible test cases and then operate inside a process where humans could verify, prioritize, patch, and ship fixes safely.
Security teams do not need vibes.
They need evidence.
A vulnerability report that comes with a working reproduction is a different artifact from a scary paragraph. It can be tested. It can be assigned. It can be fixed.
That is the line between AI slop and AI security work.
Humans still own the fix
Mozilla was explicit that it is not letting AI directly patch these bugs into production.
The system may propose patch ideas, but each real fix still goes through human engineering and review.
That is not a failure of automation.
That is maturity.
In security, the fix can be as dangerous as the bug if it breaks assumptions, adds complexity, or quietly opens a different path. A model that helps find the issue is already valuable. A model that drafts context for the fix is useful. A model that bypasses review is how you create tomorrow’s incident.
The boring discipline is the point.
The uncomfortable part
This capability helps defenders.
It also helps attackers.
Mozilla can use agentic scanning to harden Firefox. So can every serious software project. But the same basic pattern can be used to hunt for unpatched bugs in software that nobody has time to defend.
That is why this moment feels different.
Software security has always been a race between discovery and patching. AI accelerates both sides. The only sane answer is to make defensive pipelines faster, more reproducible, and more continuous before offensive pipelines become normal.
Mozilla’s lesson is not “use Mythos and relax.”
It is “build the pipeline now.”
Because the models are no longer just writing fake bug reports.
They are finding real ones.
Sources: Mozilla Hacks, TechCrunch, Anthropic Red Team