Industry case study

How Mozilla used AI to find and fix Firefox vulnerabilities

Mozilla’s Firefox team turned an early AI security evaluation into shipped browser fixes, with maintainers responsible for getting the work into a release.

Source publisherMozilla
Source published21 April 2026
Last checked

Independent Cactera analysis of publicly documented work. Cactera was not involved in this work. Company names identify the subjects, not Cactera clients or partners.

Maintainer report of shipped fixes

271 fixesfor vulnerabilities identified during the Firefox 150 AI evaluation

Read Mozilla’s account
Comparison
No controlled comparison or detection-rate denominator
Scope
Firefox vulnerabilities from the initial Mythos Preview evaluation
Timeframe
Firefox 150 release, reported 21 April 2026
The published work

The problem.

Mozilla describes uneven coverage from automated fuzzing and the time required for expert source-code review, even in a browser with established security defenses.

What changed.

The Firefox team evaluated an early Claude Mythos Preview model through its collaboration with Anthropic and worked through the vulnerabilities it identified.

As described by Mozilla.

Maintainer report of shipped fixes

What was reported.

Mozilla reports that Firefox 150 included fixes for 271 vulnerabilities from that initial evaluation. The figure concerns this release, without adding earlier Firefox findings. Source: Mozilla

What the evidence can tell us

This is Mozilla’s own release account. It does not establish a general detection rate, independent model ranking, or number of attacks prevented. A completed evaluation also cannot establish that a browser has no remaining vulnerabilities.

Cactera analysis

What we take from it.

For a product team, a security finding becomes valuable when someone can turn it into a reliable change. We would begin an AI review with the release process: who owns the affected component, how a proposed fix reaches review, and which checks must pass before it ships. Finding more issues is useful only if that route can absorb them.

The practical unit of work should be a traceable issue. Keep its affected version, reproduction evidence, proposed change and verification result together. A second engineer should be able to inspect the reasoning without rebuilding the investigation from scattered conversations. This also gives the team a useful record when a similar problem appears in another part of the application.

Coverage needs its own accounting. A long findings list can hide areas that were never meaningfully examined. We would record which components were reviewed, which assumptions the review used, and which paths still need attention. That record helps a team allocate the next assessment instead of mistaking a successful first pass for complete coverage.

Finally, remediation should protect ordinary use. A patch that removes a risky behavior may also disrupt something customers rely on. We would review the security change alongside relevant product checks, then track the release and any follow-up work. The measure of progress is an issue resolved in the supported product, with a clear owner for what remains.

A proposed method for your business

How to evaluate a similar idea.

Start with your situation and a question you can test. These are evaluation steps we would discuss before choosing an implementation.

  1. 01

    Connect review to a release

    Choose an owned component and agree who will review, test and ship any confirmed fixes.

  2. 02

    Retain the evidence

    Keep affected versions, reproduction notes and verification results with each issue so another engineer can assess it.

  3. 03

    Track unfinished coverage

    Record unexamined areas and unresolved assumptions alongside the findings that have already been addressed.

  4. 04

    Verify the delivered change

    Retest the issue, check expected product behavior and confirm that the fix reaches the supported release.

Industry case study / Source notes

Sources & credits.

Work credited to
Mozilla’s Firefox security and engineering teams
Technology / platform
Anthropic
Analysis & explanation
Cactera. Company wordmarks identify the article subjects.

Independent Cactera analysis of publicly documented work. Cactera was not involved in this work. Company names identify the subjects, not Cactera clients or partners.

A relevant next step

Bring the right question.
Let’s make it specific.

Explore how penetration testing could fit the work you have in mind.

Discuss penetration testing