Firefox 150: AI-Assisted Security, 271 Fixes, and What It Means for Business

TL;DR

• Firefox 150 adds productivity features (improved split view, multi-tab sharing, an about:translations page and a built-in PDF editor) and, crucially, patched 271 security vulnerabilities.
• Mozilla used Anthropic’s Claude Mythos Preview (an advanced AI model) to accelerate vulnerability discovery; earlier testing with Opus 4.6 found 22 bugs in Firefox 148.
• Project Glasswing coordinates vendors and AI tools so security teams can test software at scale. This helps defenders — but also raises dual-use and governance risks.
• CISOs should pilot AI agents with strict controls: logging, human triage, prompt governance, and legal review before rolling out broadly.

What’s new in Firefox 150

Firefox 150 focuses on two things: making the browser a better workspace and hardening it. On the productivity side you’ll find a smoother split view, more flexible tab management and sharing (some users reported the multi-tab Share option behaving inconsistently), a new about:translations page for in-browser translations, and an upgraded PDF editor that lets you reorder, delete and export pages without leaving the tab.

On the security side, Mozilla reported that it patched 271 vulnerabilities in this release. That scale of fixes is unusual for a single browser update and is the headline most businesses should notice.

Why AI-assisted security (AI agents) changed Firefox 150

Mozilla used an Anthropic model called Claude Mythos Preview—an advanced AI system tuned to help with code and security testing—to accelerate discovery and remediation. That work followed earlier testing with Anthropic’s Opus 4.6, which Mozilla says helped find 22 bugs in Firefox 148.

Mozilla has been using advanced AI models since February to root out hidden vulnerabilities in the browser.

Put simply: automated AI tools found many candidate issues, and humans reviewed and fixed them. For defenders, that’s a big deal. Attackers only need one critical vulnerability to exploit; defenders need to find them all. AI agents can sweep code and interfaces much faster than people alone, narrowing that gap.

Project Glasswing: vendors pairing AI with security teams

Project Glasswing is a multi-vendor effort that pairs advanced AI models with security teams to scale defensive testing. Anthropic is a key partner, and major vendors including Apple, Google and Microsoft are participating. The project shows the industry is experimenting with AI agents as standard tools for vulnerability discovery, not just a niche research trick.

That collaboration matters for businesses: it signals a shift from manual, limited testing toward automated, model-enabled testing pipelines run by vendors. Expect more frequent, broader sweeps of software components as part of normal release cycles.

How these AI models actually find bugs

Here’s how the process typically works, at a high level:

• Automated fuzz testing: the model generates lots of unexpected inputs to APIs, UI fields, and file parsers to see what breaks.
• Code-path exploration: models suggest sequences of actions or inputs that reach deep, unusual parts of the codebase.
• Hypothesis generation: the model proposes likely vulnerability types and points to the lines or components to inspect.

Think of AI as a high-speed metal detector for software: it pings a large area and flags many signals. Humans still need to dig and confirm whether each signal is treasure or scrap. That combination—speed from AI, judgment from engineers—creates coverage you couldn’t scale easily before.

Concrete (hypothetical) example: an AI scan targets the new translation feature. It feeds malformed translation files and finds a path that causes an out-of-bounds read in a parser used by the translation system. The model flags the sequence, a human triage team reproduces it, and engineers ship a patch that closes the vulnerability.

Dual-use risk: attackers get smarter too

The same AI techniques that help vendors can help attackers. Models that speed discovery can be retooled to craft exploit chains, automate reconnaissance, or find zero-days faster than traditional tools.

This creates three practical risks for businesses:

• Faster offensive discovery means shorter windows for vendors to patch.
• Model outputs can include dangerous instructions—so controls are needed to prevent misuse.
• False positives and low-signal findings can swamp teams if triage isn’t scaled properly.

Defenders must assume adversaries will adopt AI agents. That changes priorities: invest not only in detection and patching, but in rapid verification, hardened defaults, and reducing attack surface proactively.

Operational checklist for CISOs and engineering leaders

1. Pilot in a controlled environment. Run AI-assisted scans against staging systems only until you validate output quality and safety controls.
2. Require human triage for every AI finding. Use experienced engineers to reproduce, validate, and score impact before any remediation work starts.
3. Log everything. Record prompts, model outputs, who ran scans, and action taken. Audit trails matter for compliance and post-incident review.
4. Restrict prompt and output capabilities. Block prompts that ask for exploit code and prevent models from returning executable payloads in clear text.
5. Integrate with patch management. Feed validated findings into your bug tracker and SLAs so fixes get prioritized and rolled out quickly.
6. Review legal and privacy risks. Check whether code or data sent to third-party models violates contracts or data protection rules; prefer on-prem or private-cloud model hosting when necessary.
7. Measure ROI and hidden costs. Track time-to-fix improvements, false-positive rates, and the engineering hours needed for triage to justify scale-up decisions.

Questions to ask vendors before adopting AI-assisted testing

• Can scans run on-prem or in a private cloud?
• Do you keep logs of prompts and outputs, and who can access them?
• What controls prevent exploit-generation or dangerous payloads?
• How do you measure false positives and tune the model over time?
• Who is liable if a scan discloses sensitive data?

Strategic implications and ROI

Adopting AI-assisted vulnerability discovery can cut time-to-find and time-to-fix dramatically. Faster patch cycles reduce breach windows and can decrease incident costs. Vendors that use these tools responsibly may win trust from enterprise customers because they deliver broader, faster coverage.

There are hidden costs: governance, integration work, human triage labor, and legal reviews. A realistic ROI assessment should balance faster remediation and reduced incident risk against the engineering hours for verification and the governance overhead required to run models safely.

Key takeaways

• Firefox 150 combined user-facing productivity features with an unusually large security cleanup: 271 patched vulnerabilities.
• Mozilla credited Anthropic’s Claude Mythos Preview (and prior Opus testing) for accelerating vulnerability discovery.
• Project Glasswing shows major vendors are experimenting with AI agents to scale defensive testing.
• AI agents can equalize defenders’ capabilities—but they create governance, legal and operational risks that teams must manage.
• Start small, log everything, require human verification, and integrate validated findings into existing patch pipelines.

Frequently asked questions

• How many vulnerabilities did Firefox 150 fix?

  Mozilla reported that Firefox 150 patched 271 vulnerabilities.

• How were so many bugs found quickly?

  Mozilla used Anthropic’s Claude Mythos Preview—an advanced AI model—to help surface candidate issues. Earlier testing with Opus 4.6 reportedly found 22 bugs in Firefox 148.

• Are AI-generated findings reliable enough to auto-patch?

  Not yet. Models produce useful leads but require human triage to confirm impact, rule out false positives, and ensure fixes are safe.

• Should defenders be worried attackers will use the same AI techniques?

  Yes. The dual-use nature of these tools means defenders should assume adversaries will adopt similar capabilities and adjust defenses accordingly.

• How should organizations start adopting AI-assisted security?

  Pilot in staging, log prompts and outputs, enforce strict prompt governance, require human verification, and integrate validated findings into existing patch workflows.

Firefox 150 is a clear signal: AI agents are moving from experiments into vendor toolchains for vulnerability discovery. For business leaders, the sensible path is proactive adoption—carefully governed and human-supervised—so your team gets the benefit of speed without taking on new, unmanaged risk.