AI security bug detection has crossed a critical threshold. Mozilla just patched 423 Firefox security vulnerabilities in a single month using Anthropic’s Mythos Preview and other AI models, demonstrating that frontier AI systems now match the capabilities of elite human security researchers—while raising uncomfortable questions about what else these systems can do.
Key Takeaways
- Mozilla fixed 423 Firefox security bugs in one month using AI security bug detection tools, primarily Mythos Preview.
- Firefox 150 (released May 2026) included 271 vulnerabilities identified by Mythos Preview alone—over 12 times more than Claude Opus 4.6 found in February 2026.
- Mythos Preview operates under restricted access via Anthropic’s Project Glasswing, distributed to only a handful of organizations.
- Mozilla’s CTO stated the defects are “finite” and defenders can “finally find them all.”
- UK AI Security Institute confirmed Mythos can execute autonomous multi-stage network attacks, chaining small vulnerabilities into devastating exploits.
How AI security bug detection scaled vulnerability patching
The numbers alone tell the story. In February 2026, Claude Opus 4.6 scanned nearly 6,000 C++ files in Firefox’s codebase and produced 112 unique reports, with 22 confirmed as security-sensitive bugs that made it into Firefox 148. That was already exceptional—22 bugs in a single month exceeded any monthly total from 2025. Then Mythos Preview arrived. Firefox 150, released in May 2026, included fixes for 271 vulnerabilities identified by Mythos Preview. The jump from 22 to 271 represents a 12-fold increase in detection speed and scope.
What made this acceleration possible? Mozilla’s security team received bug reports from Anthropic’s frontier red team with minimal test cases included—a crucial difference from typical AI-generated submissions, which often flood researchers with low-quality findings that require extensive manual triage. The minimal test case format allowed Mozilla’s team to verify findings faster than they could with traditional fuzzer output or crowdsourced submissions. This efficiency is what enabled the 423-bug total across all AI models in the month.
AI security bug detection versus human elite researchers
Mozilla’s own assessment is striking. In a blog post titled “The zero-days are numbered,” the Firefox team stated: “We haven’t seen any bugs that couldn’t have been found by an elite human researcher.” This is not hyperbole masquerading as humility—it is a direct claim that AI security bug detection has reached parity with the world’s best humans, not exceeded them.
The key difference is speed and scale. Elite human researchers find bugs by reasoning through source code—a bottleneck that has constrained security work for decades. Computers were completely incapable of this reasoning task a few months ago. Now they excel at it. Mozilla’s team, with years of experience analyzing the work of world-class security researchers, found that Mythos Preview matched their capabilities across every category and complexity of vulnerability. No gap exists. The model finds what humans find, but faster and at scale.
Traditional fuzzers—automated tools that bombard software with random inputs to trigger crashes—cannot find these bugs. Fuzzers excel at certain classes of defects but fail at vulnerabilities that require reasoning about code logic rather than brute-force input generation. This is where AI security bug detection gains its edge: it thinks through code the way elite researchers do, but without the fatigue or time constraints.
The dual-use threat hiding in the capability
Mythos Preview’s power cuts both ways. The UK AI Security Institute confirmed that Mythos can execute autonomous multi-stage network attacks. It chains multiple small vulnerabilities together into devastating exploits, reconstructs source code from deployed software, and builds custom tools for lateral movement and data extraction. These are capabilities that previously required teams of expert attackers working in coordination.
The irony is bitter. The same AI security bug detection capability that allows Mozilla to patch vulnerabilities at scale also enables attackers to find and weaponize them autonomously. A system that can reason through code to find security flaws can reason through code to exploit them. Anthropic is distributing Mythos Preview under Project Glasswing, a restricted access programme given to only a handful of organizations, precisely because of this dual-use risk.
That caution proved insufficient. On the day Anthropic announced Glasswing, unauthorized users accessed Mythos Preview by guessing its URL via a third-party vendor. Anthropic is investigating the breach. The incident underscores a hard truth: frontier AI security capabilities cannot be contained through access restrictions alone. Once the capability exists, containment is a temporary measure, not a permanent solution.
What this means for the future of vulnerability research
Mozilla’s CTO Bobby Holley framed the moment with unusual optimism. He stated that defects are “finite” and that defenders can “finally find them all.” The implication is that AI security bug detection removes the human bottleneck that has allowed vulnerabilities to hide in plain sight for years. If defenders can find them all, attackers cannot rely on the assumption that their exploits will remain undiscovered.
But this assumes defenders get there first. The same reasoning applies in reverse: if AI can find all vulnerabilities, attackers using the same AI will find them too. The race is on. Organizations that adopt AI security bug detection early gain a window to patch before adversaries weaponize the same flaws. That window will close as the capability diffuses.
The broader shift is philosophical. For decades, security research was a human-constrained discipline. You needed elite researchers, and they were scarce. Now it is a capability-constrained discipline. You need frontier AI access, and that access is controlled but spreading. Anthropic’s decision to partner with Mozilla signals that the company views defensive security as a legitimate use case for Mythos Preview. But the same model, in different hands, becomes an offensive weapon. That tension will define AI security for the next several years.
Is Mythos Preview available to the public?
No. Mythos Preview is an unreleased frontier model distributed exclusively under Anthropic’s Project Glasswing programme to a small number of organizations. Mozilla is one of them, but public access is not planned. The restriction reflects Anthropic’s concern about dual-use risks and autonomous attack capabilities.
How many Firefox bugs did Mythos Preview find compared to earlier models?
Mythos Preview identified 271 vulnerabilities in Firefox 150 alone. That is over 12 times the 22 bugs Claude Opus 4.6 found in February 2026, demonstrating a dramatic leap in detection capability within just a few months.
Can AI security bug detection find bugs that human researchers cannot?
No. Mozilla explicitly stated that it has not found any bugs via AI security bug detection that could not have been found by an elite human researcher. The advantage is speed and scalability, not the discovery of a new category of vulnerability. Mythos matches human elite capability but applies it at machine scale.
The story of Mozilla’s 423 Firefox security fixes is not about AI surpassing humans. It is about AI finally matching them—and what happens when that matching capability spreads to attackers. The window for defenders to act is open, but it is closing. Organizations that move now will patch their vulnerabilities before adversaries weaponize the same AI-powered reasoning. Those that wait will face an adversary that thinks like their best researchers and never sleeps.
This article was written with AI assistance and editorially reviewed.
Source: TechRadar
