Mozilla fixed an unprecedented 423 Firefox security vulnerabilities during April 2026 — nearly 19 times the browser's historical monthly average of roughly 22 fixes — using an agentic AI security pipeline built around Claude Mythos Preview, Anthropic's flagship large language model. The effort, publicly disclosed in a Mozilla Hacks blog post on May 7, 2026, produced Firefox 150's security advisory MFSA 2026-30 listing more than 41 CVEs, including ten rated High severity. There is no known active exploitation of any of these vulnerabilities. Every Firefox and Thunderbird user below version 150 should patch now.
Firefox 150 Vulnerabilities: Technical Breakdown
The Firefox 150 release patches a broad range of vulnerability classes across the browser's most security-sensitive subsystems. Of the 271 bugs attributed to Claude Mythos Preview, the majority fell into three rollup CVEs designed to group related memory safety failures discovered through the AI pipeline:
- CVE-2026-6784 (High severity): Covers 154 memory safety bugs fixed in Firefox 150 and Thunderbird 150. These bugs showed evidence of memory corruption in Firefox 149 that Mozilla assessed as potentially exploitable to run arbitrary code.
- CVE-2026-6785 (High severity): Covers 55 memory safety bugs spanning Firefox ESR 115.35, ESR 140.10, Thunderbird ESR 140.10, and Firefox/Thunderbird 150.
- CVE-2026-6786 (High severity): Covers 107 bugs across Firefox ESR 140.10, Thunderbird ESR 140.10, and Firefox/Thunderbird 150.
Beyond the rollup CVEs, MFSA 2026-30 lists ten individual High-severity CVEs targeting specific browser components:
| CVE | Component | Bug Class | |—–|———–|———–| | CVE-2026-6746 | DOM Core & HTML | Use-after-free | | CVE-2026-6747 | WebRTC | Use-after-free | | CVE-2026-6748 | Web Codecs | Uninitialized memory | | CVE-2026-6749 | Canvas2D | Information disclosure via uninitialized memory | | CVE-2026-6750 | WebRender | Privilege escalation | | CVE-2026-6751 | Web Codecs | Uninitialized memory | | CVE-2026-6752 | WebRTC | Incorrect boundary conditions | | CVE-2026-6753 | WebRTC | Incorrect boundary conditions | | CVE-2026-6754 | JavaScript Engine | Use-after-free | | CVE-2026-8091 | Audio/Video Playback | Incorrect boundary conditions |
Three CVEs — CVE-2026-6746, CVE-2026-6757, and CVE-2026-6758 — are explicitly credited to researchers "using Claude from Anthropic" in the advisory. Mozilla does not publish CVSS v3 scores (a standardized 0–10 numeric rating of vulnerability severity) in its advisories, instead using an internal High/Moderate/Low scale. High maps roughly to CVSS 7.0–10.0: remotely exploitable or locally exploitable with serious consequences.
A use-after-free (UAF) vulnerability occurs when a program continues to use a pointer to memory after that memory has been freed and potentially reallocated to a different object. In a browser renderer, a UAF can allow a malicious web page to corrupt internal state and, in the worst case, achieve remote code execution (RCE — the attacker runs arbitrary commands on the victim's machine) simply by loading a specially crafted page. Uninitialized memory bugs leak data from one security boundary to another, potentially exposing cryptographic material or process memory across origin boundaries.
How the AI Agentic Pipeline Works
Mozilla's security team did not simply feed Firefox source code into an AI chatbot. They built a custom agentic harness — an automated orchestration system that instructs an AI model to create and execute reproducible test cases, dynamically validate hypotheses about potential code paths, and iterate on findings without human involvement between cycles.
The pipeline attacks Firefox from three complementary angles simultaneously:
- Fuzzing — automated tools that generate massive volumes of malformed or unexpected inputs to trigger crashes and expose memory errors.
- Manual code review — traditional human-led inspection of security-critical components.
- AI agentic pipeline — the new addition, running Claude Mythos Preview and, in earlier experiments, Claude Opus 4.6, GPT-4, and Claude Sonnet 3.5.
The combination surfaced vulnerabilities that conventional fuzzing routinely misses. Fuzzing excels at finding surface-level input validation failures, but struggles with complex state-dependent bugs involving multiple subsystems interacting over time. Among the most striking finds: a 15-year-old bug in Firefox's <legend> element handling (internally tracked as bug 2024437) and a 20-year-old heap use-after-free in the XSLTProcessor DOM API (bug 2025977) — a bug where a web page could trigger memory corruption through XSLT (a language used to transform XML documents) without any user interaction beyond loading the page.
Multiple sandbox escape vulnerabilities were also identified. A sandbox escape is a bug that allows code running inside Firefox's restricted renderer process — the isolated container that handles untrusted web content — to break out into the host operating system. Sandbox escapes typically need to be chained with a second vulnerability to achieve full-system compromise, but they represent a critical attack surface in any layered browser defense.
Exploitation Status and Threat Landscape
None of the 423 bugs fixed in Firefox 150 are known to be under active exploitation as of May 8, 2026. Mozilla found and patched these vulnerabilities before any threat actor weaponized them. CISA has not added any CVEs from MFSA 2026-30 to its Known Exploited Vulnerabilities (KEV) catalog — CISA's authoritative list of vulnerabilities confirmed to be actively exploited in real-world attacks.
This proactive posture is the intended outcome: the AI pipeline discovered latent vulnerabilities that would likely have persisted undetected for months or years. For defenders, that means this patch cycle is routine maintenance, not emergency response. Prioritize it within your normal patching cadence rather than treating it as a crisis.
Who Is Affected
Any user, organization, or embedded deployment running Firefox or Thunderbird below version 150 is affected by at least some of the 41+ CVEs in MFSA 2026-30:
- Desktop Firefox (Windows, macOS, Linux) below version 150
- Firefox ESR (Extended Support Release — a stable branch maintained for enterprise use) below version 140.10 or 115.35
- Thunderbird below version 150 or ESR 140.10
- Chromium-based browsers (Chrome, Edge, Brave) are unaffected — these vulnerabilities are specific to Firefox's Gecko rendering engine and SpiderMonkey JavaScript runtime
Enterprise deployments running Firefox ESR for stability should prioritize the upgrade to ESR 140.10. The high concentration of use-after-free and uninitialized memory bugs in WebRTC (a browser API for real-time audio and video communication) and JavaScript engine subsystems is directly relevant to any environment where users open untrusted web content — which is effectively every corporate browsing environment.
What You Should Do Right Now
- Update Firefox to version 150 via
Help → About Firefox. The browser auto-updates if background updates are enabled; force a check if you haven't restarted Firefox recently. Download directly from the official Firefox release page if needed.
- Update Thunderbird to version 150 or ESR 140.10 for any deployments running the Mozilla email client.
- Verify version across your fleet before declaring compliance:
# Linux: check installed Firefox version
firefox --version
# macOS: check Firefox version
/Applications/Firefox.app/Contents/MacOS/firefox --version
# Windows (PowerShell): check Firefox version
(Get-Item "C:\Program Files\Mozilla Firefox\firefox.exe").VersionInfo.FileVersion
- Audit for unmanaged installs. Copies of Firefox installed outside your software inventory — by individual users on personal laptops, for instance — are a common blind spot. Query your EDR (Endpoint Detection and Response — security tooling that monitors endpoint activity) or vulnerability scanner for unpatched Firefox instances across your estate.
- Review WebRTC exposure. Several High-severity CVEs (CVE-2026-6747, CVE-2026-6752, CVE-2026-6753) target WebRTC specifically. If your organization enforces browser policies that disable WebRTC for compliance or privacy reasons, verify those policies are active and applied.
- Monitor CISA KEV for any future additions from this advisory batch. If any MFSA 2026-30 CVE appears on the KEV catalog, treat patching as an emergency and escalate your timeline accordingly.
Background: Understanding the Risk
This release marks a meaningful inflection point in browser security methodology, not just a large patch batch. The previous record for Firefox monthly vulnerability fixes was around 76 in March 2026. April's 423 — achieved partly because February's initial Anthropic red team engagement revealed the pipeline's potential — represents a qualitative shift in how proactive browser hardening can scale.
The specific bug classes uncovered illustrate why this matters. Memory safety bugs in rendering engines are the foundational building block of browser exploitation chains. High-profile browser zero-days exploited in the wild — used by nation-state actors and commercial spyware vendors alike — most commonly begin with exactly this type of use-after-free or uninitialized memory read in a DOM, graphics, or JavaScript component. The fact that a 20-year-old UAF in XSLT processing survived until 2026 across hundreds of millions of Firefox installations is a sobering demonstration of how poorly conventional security tools scale to large, complex codebases.
CVE-2026-6749, the information disclosure bug in the Canvas2D API, deserves particular attention. The Canvas2D API is used by virtually every modern web application for graphics rendering, image manipulation, and (controversially) browser fingerprinting. A memory leak in Canvas2D could expose renderer process memory to a compromised or malicious web page, potentially enabling cross-origin data theft or side-channel attacks against other browser tabs.
For the wider security industry, this release validates the premise that agentic AI pipelines — systems where an AI model takes multi-step actions autonomously rather than just answering questions — can find high-severity vulnerabilities in production software at a scale and speed that human review teams and conventional fuzzers cannot match alone.
The Attribution Debate: Model or Harness?
Not everyone is convinced Mythos itself deserves the credit. Security researcher Davi Ottenheimer challenged Mozilla's methodology publicly, noting that the browser maker never benchmarked Claude Opus 4.6 — a substantially less expensive Anthropic model available before Mythos — independently against the same codebase before attributing 271 discoveries to Mythos alone.
Ottenheimer's own experiments, using cheaper models (Claude Sonnet 4.6, Claude Haiku 4.5) paired with a similar agentic harness, reportedly surfaced eight Firefox vulnerabilities in two minutes for under one dollar in API cost — two of which matched Mythos discoveries. His conclusion: "The improvement may be entirely in the harness, not as much in the model."
Mozilla's team acknowledged "dramatically improved techniques for harnessing these models" in the same blog post, which lends weight to the argument that the orchestration layer — how prompts are structured, how test cases are generated, how results are triaged and validated — may matter as much as raw model capability.
For organizations evaluating AI-assisted security tooling investments, this distinction has real commercial implications. If a well-designed harness achieves comparable coverage with a cheaper model, the defensible competitive advantage lies in harness design, not in access to a specific frontier model. Regardless of attribution, the outcome is not in dispute: 271 latent Firefox vulnerabilities that could have been exploited for years were found and patched before a single user was harmed.
Conclusion
Firefox 150 patches 41+ CVEs — including 316 grouped memory safety bugs and ten individual High-severity vulnerabilities in DOM processing, WebRTC, WebRender, and the JavaScript engine — discovered through Mozilla's AI-assisted agentic security pipeline. No active exploitation is known. All Firefox and Thunderbird users should update to version 150 immediately; enterprise administrators running ESR should deploy ESR 140.10 within this week's patching cycle.
For any query contact us at contact@cipherssecurity.com
