Firefox 150: Mythos AI Finds 271 Zero-Days in Paradigm Shift

Claude Mythos AI discovered 271 zero-days in Firefox 150. Learn why this marks a paradigm shift in cybersecurity and what it means for defenders.

Firefox 150: Mythos AI Finds 271 Zero-Days, A Paradigm Shift

Mozilla has released Firefox 150 with fixes for 271 zero-day vulnerabilities identified by Claude Mythos Preview during a single evaluation in the week of April 15, 2026. The intervention marks a significant shift in how defenders can approach vulnerability discovery, reducing time and costs compared to traditional human expertise.

The context: the Mozilla-Anthropic collaboration from February 2026

Since February 2026, the Firefox team has been collaborating with Anthropic to test the browser using frontier AI models. The agreement is part of Project Glasswing, the controlled access program for Mythos announced on April 7, 2026. The partnership already produced concrete results with Firefox 148, when Claude Opus 4.6 identified 22 security-sensitive bugs, 14 of which were classified as high-severity.

To understand the scale of the change, just compare the numbers: in 2025, Mozilla resolved approximately 73 high-severity vulnerabilities for the entire year. Mythos found 271 in a single evaluation on Firefox. The model scored 93.9% on SWE-bench and 97.6% on USAMO, demonstrating advanced reasoning capabilities applied to code security.

The technical capabilities of Mythos: from discovery to exploit

Within the Firefox JavaScript shell, Mythos converted 72.4% of the identified vulnerabilities into working exploits, gaining register control in 11.6% of attempts. On the Firefox test set, it produced 181 working exploits and gained register control in 29 cases. For comparison, Opus 4.6 on the same testbed only managed to produce 2 successful exploits.

According to data provided by Mozilla, the cost for 1,000 OpenBSD scans that found "several dozen" findings was less than $20,000, while a full n-day exploit reproduction cost less than $1,000. These numbers indicate a drastic reduction in costs compared to traditional security research methods.

The UK AI Security Institute evaluated Mythos as capable of autonomously executing multi-stage network attacks. In "The Last Ones," a 32-step corporate network attack simulation, the model completed the attack in 3 out of 10 attempts.

Historical bugs: 27 years in OpenBSD, 17 in FreeBSD

Claude Mythos found vulnerabilities that had eluded human researchers for decades. In OpenBSD, it identified a 27-year-old TCP bug. In FFmpeg, it discovered an H.264 vulnerability dating back 16 years. In FreeBSD, it identified an NFS RPC flaw (CVE-2026-4747) that was 17 years old.

These findings demonstrate that AI is not limited to replicating known patterns but can analyze complex codebases with a systematicity that the human approach struggles to match in scale. As the Firefox team wrote, citing the need to "shake off the vertigo and get to work," the numbers demand a reflection on security methodologies.

The disclosure wave: over 1,000 vulnerabilities, 99% unpatched

The 271 Firefox vulnerabilities represent only a fraction of Mythos's overall work. The model has found over 1,000 total vulnerabilities across Firefox, Chrome, Linux, FreeBSD, OpenBSD, and cryptographic libraries. Of these, over 99% remain unpatched.

Anthropic has adopted a 90+45 day responsible disclosure timeline. In the time window between disclosure and the release of patches, exposed systems remain potentially vulnerable. Firefox was the first major target to complete the fix cycle, with the Android patch dated April 15, 2026.

The change of perspective according to Mozilla and Bruce Schneier

In a post published Tuesday, Mozilla wrote that "defenders finally have a chance to win, decisively." However, the team also warned: "For a hardened target, just one such bug would have been red-alert in 2025, and so many at once makes you stop to wonder whether it's even possible to keep up."

The central question is whether the technology favors attackers or defenders. Bruce Schneier commented: "Assuming the defenders can patch, and push those patches out to users quickly, this technology favors the defenders." The key lies in the speed of response: finding vulnerabilities is only the first step; the ability to distribute patches determines the actual impact.

Mozilla also specified: "We haven't seen any bugs that couldn't have been found by an elite human researcher." The difference is not in the quality of the findings, but in the scale and costs. What previously required years of human expertise has now become scalable through intelligent automation.

The implications for the security ecosystem

The traditional cybersecurity paradigm saw attackers as having the advantage: they only needed to find a single bug to compromise a system. Defenders, meanwhile, had to protect every possible entry point. Mythos reverses this dynamic by making systematic vulnerability discovery a routine, high-volume operation.

The real challenge over the next 90-135 days will involve managing the disclosure wave. While Firefox has completed its patch cycle, the more than 1,000 vulnerabilities identified in Linux, BSD, FFmpeg, and other platforms are still awaiting fixes. This time window creates a period of high risk for systems that have not yet been updated.

The economic aspect is equally relevant. With costs below $1,000 for a full exploit reproduction, the barrier to entry for security testing is dropping drastically. This makes in-depth security audits accessible even to organizations with limited budgets, but simultaneously democratizes potentially dual-use tools.

Frequently asked questions

What is Claude Mythos Preview?: Claude Mythos Preview is a frontier AI model from Anthropic announced on April 7, 2026, accessible through the Project Glasswing program. It scored 93.9% on SWE-bench and 97.6% on USAMO, demonstrating advanced capabilities in code analysis and vulnerability discovery.
How many vulnerabilities did Mythos find in Firefox?: Mythos identified 271 zero-day vulnerabilities in Firefox 150 during a single evaluation. Previously, Claude Opus 4.6 had found 22 in Firefox 148, 14 of which were classified as high-severity.
What historical bugs did Mythos discover?: The model identified a 27-year-old TCP bug in OpenBSD, an H.264 vulnerability in FFmpeg dating back 16 years, and an NFS RPC flaw in FreeBSD (CVE-2026-4747) that was 17 years old.
How many total vulnerabilities did Mythos find?: Over 1,000 total vulnerabilities across Firefox, Chrome, Linux, FreeBSD, OpenBSD, and cryptographic libraries. Of these, over 99% have not yet been patched.
What is Anthropic's responsible disclosure timeline?: Anthropic adopts a 90+45 day timeline for responsible disclosure, giving vendors time to develop and distribute patches before vulnerability details are published.

This article is a summary based exclusively on the listed sources.