Anthropic’s preview of Mythos now features a scoreboard showcasing the real-world achievements of companies involved in Project Glasswing, which have spent several weeks working with the model.
The data emerging from these early deployments paints a startling picture of a security landscape that has been completely accelerated by automation.
Mozilla revealed that before deploying Anthropic’s Mythos, Firefox typically fixed around 20–30 security bugs per month. After introducing the model in early 2026, the number of fixes rose to 76 in March and 423 in April, with 271 of those vulnerabilities directly linked to Mythos-assisted discovery.
It helped the company find numerous critical issues and exposed historical legacy bugs, including a 15-year-old flaw in an HTML element and a 20-year-old XSLT vulnerability.
Staying on top of this unprecedented volume of bugs “led to a lot of work and long days over the last few months,” Mozilla noted, adding that over 100 people contributed code to this effort to ship the most secure Firefox yet.
The company also said it felt a sense of ‘vertigo’ when the findings first came into focus. “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,” wrote Mozilla in a blog post. It expects many more companies defending bugs to feel the same way.
Besides Mozilla, the UK’s AI Security Institute (AISI) reported that Mythos is the first model to solve both of its highly complex cyber ranges.
A newer checkpoint of the model successfully navigated “The Last Ones” (a 32-step simulated corporate network attack) in 6 out of 10 attempts, and became the first ever to crack the previously unsolved “Cooling Tower” range, succeeding in 3 out of 10 runs.
Palo Alto Networks said that after scanning 130+ products, including models such as Claude Mythos and GPT-5.5-Cyber, it patched 26 Common Vulnerabilities and Exposures (CVEs) covering 75 issues—far above its usual monthly volume of fewer than 5 CVEs.
In Anthropic’s testing across 1,000 open-source projects, Mythos found 6,202 high- or critical-severity vulnerabilities and 23,019 total bugs.
The Talent Paradox & The Fallacy of Patching Faster
John Scimone, the Chief Security Officer of Dell Technologies, told AIM that the industry’s longstanding talent shortage is being offset by models like Mythos, which accelerate vulnerability discovery.
“As somebody who's built and led cybersecurity teams for decades, the number one challenge I've had, regardless of whether it has been attracting, recruiting, and retaining qualified talent,” says Scimone.
He argues AI is a force multiplier despite the surge in discovered vulnerabilities that may increase the workload on humans.
“Potentially you need more people to then understand the issues [found by the model] and resolve them — but when you look at the big picture, the net result is that you can produce cybersecurity outcomes far more efficiently with less dependency on that expert labour, which can produce far greater scale and impact.”
Scimone advocates fighting fire with fire and believes that any human-expert challenge can also be solved through extended use by the models themselves, pointing to the high percentage of code generated by AI coding models.
Besides suggesting that Mythos will change how human experts work, it also raises the need to fundamentally rewire how human experts have worked and built their workflows.
In an interaction with AIM, Keshava Murthy, co-founder and CEO of Matters.ai, an AI security firm, says there needs to be a fundamentally different approach to building patching workflows and that companies cannot hire their way out of the problem.
“When a vulnerability can be turned into a working exploit in under three days, traditional cycles of scanning, scoring, scheduling, and patching consume time enterprises simply do not have,” says Murthy. “Even well-resourced teams cannot sprint on every finding simultaneously, forcing them to make difficult choices about what to fix first.”
Similarly, Cloudflare shared how the instinct to simply “patch faster” became a dangerous trap in their experience with Mythos.
While the model excelled at complex exploit chain construction and automated proof generation, Cloudflare warned that compressing the response cycle to meet extreme SLAs, like two hours from disclosure to patch, inevitably requires skipping critical regression testing.
“The bugs you ship when you skip regression testing tend to be worse than the bugs you were trying to patch,” the company noted in its blog post.
Cloudflare found that pointing a generic coding agent at a repository fails; instead, deploying a specialised, multi-stage “harness” in which multiple AI agents work within narrow scopes and cross-validate findings drastically reduced triage noise.
Rather than chasing every alert across the board, Murthy adds that companies should reassess their standard severity scores, as many lack business context.
“A critical CVE in an isolated system holding no sensitive data is practically less urgent than a medium CVE in a core system processing healthcare data,” says Murthy. “By routing incoming vulnerability signals through a real-time map of where sensitive data actually lives, it automatically prioritises the alerts that pose a genuine business risk.”
He adds that it isn’t about patching faster across the board, but about being ‘ruthlessly’ accurate about which ten things actually matter right now, and securing those before the window closes.
Ultimately, organisations must shift toward application-level architectural changes that block exploit paths by default, making exploitation harder even when vulnerabilities remain unpatched.
Mythos Unmarks the True Cost of Tech Debt
Besides the human bottleneck that Mythos may introduce, Mozilla also stated that it helped fix legacy bugs on its platform.
“The emergence of Mythos changes the tech debt conversation completely,” tells Vrajesh Bhavsar, CEO of Operant AI, a runtime cybersecurity platform for agentic AI systems, in an interaction with AIM.
For decades, enterprises assumed that if a legacy bug survived years of audits, it was hidden well enough to be safe. Bhavsar notes that Mythos destroys that assumption entirely, proving that twenty years of tech debt can be mapped and weaponised at machine speed.
Similarly, Murthy argues that legacy systems were frequently reviewed by genuinely skilled people working under human constraints.
“Human reviewers face time limits and a natural tendency to trust long-running, stable software,” he says. “AI carries no such assumptions, reading millions of code lines with consistent attention and finding complex patterns across subsystems that individual reviewers might miss.”
For many enterprises, the legacy risk profile is much higher than assumed, states Murthy, and the decision to leave older systems alone usually rests on the belief that they are stable.
“Enterprises must determine which legacy architectures have actual, traceable routes to sensitive data through APIs or shared credentials,” he says, adding that a structural flaw in an authentication layer safeguarding customer records is regarded as a different category of problem compared to the same flaw in an isolated reporting tool. Securing technical debt requires treating it as a tiered risk based solely on what each system can access.
The Move to Runtime Defence
The fundamental issue, Bhavsar warns, is that traditional ‘identify and patch’ playbooks completely break down against these new models.
“Every security tool most organisations rely on was built before emergent behaviour was a proven threat,” he explains. “They were designed to catch known threats, moving at human speed, exploiting anticipated weaknesses.”
Because AI capabilities now rely on organic learning rather than explicit programming, the threat landscape has fundamentally shifted. “Emergent behaviour is none of those things, it is unexpected, it operates at machine speed, and by definition it has not been seen before,” Bhavsar points out.
Instead of treating foundational tech debt as a slow-burning risk for a future roadmap, Bhavsar states that companies must accept it as an immediate, active exposure.
Cloudflare also found that Mythos exhibited inconsistent organic refusals, where equivalent exploit-generation tasks sometimes succeeded and sometimes failed depending on framing or execution context, suggesting that emergent safety behaviours alone are insufficient as a reliable security boundary.
All things considered, Mythos will now force organisations to rethink their cybersecurity infrastructure and how teams operate, both in preparing to fix bugs it identifies and in anticipating how attackers might use models like these once they eventually land in their hands.
“The Mozilla situation makes the problem viscerally concrete, and Mozilla is one of the most technically sophisticated organisations in the world. They had the engineers, the culture, and the infrastructure. They still got overwhelmed,” says Bhavsar.
“Enterprises cannot hire their way out of this problem,” he says.
Bhavsar explained that for enterprises without 100 developers on standby, updating their patching cycle is not the answer. “The patching model itself is broken,” he adds. He advocates for stopping reliance on patching as the primary defence and shifting to active runtime defence, detecting and blocking anomalous AI behaviour as it occurs, not after.
“That should be a layer that does not wait for a threat to be named, catalogued, or patched, but watches behaviour as it happens and blocks anomalies the moment they emerge.”
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.