Claude Mythos Can Hack Networks Alone, UK Safety Lab Warns

The UK's Artificial Intelligence Security Institute has flagged Anthropic's Claude Mythos as the first AI system capable of autonomously executing multi-step network intrusions. The government-backed research lab found that these attacks would take a skilled human professional several days to execute by hand, according to Yahoo News.

Anthropic announced Mythos last week and immediately withheld it from public release, citing the severity of what it had built. The company's own language was unusually stark: Mythos could "reshape cybersecurity," with potential impacts on economies, public safety, and national security that Anthropic itself described as "severe."

The AISI's independent assessment added historical perspective. The institute noted that, as recently as two years ago, frontier AI models could barely clear introductory-level cyber challenges. Mythos represents what the AISI called a "step up" in that trajectory, and the institute warned that more capable systems are already in development.

What Mythos actually does

Mythos has surfaced thousands of previously unknown security vulnerabilities across widely-used web browsers and operating systems. More consequential is what comes next: the model exploits those weaknesses in coordinated sequences, chaining multiple steps into attack campaigns that operate without human prompting at each stage. That end-to-end autonomy separates it from the AI-assisted hacking tools that researchers have discussed for the past two years.

Rather than a commercial launch, Anthropic distributed a limited version of Mythos to more than 40 organizations. Yahoo News confirmed that recipients include JP Morgan, Google, and Nvidia. The stated rationale is defensive: giving well-resourced institutions early visibility into the class of attacks Mythos enables so they can build countermeasures. The decision to simultaneously characterize the model as a potential national security threat and distribute it to a curated corporate list will invite scrutiny from regulators trying to establish what responsible disclosure means for offensive AI.

The AISI's warning lands against a backdrop of measurably shifting expectations. A Just Capital spring 2026 survey found that 52% of corporate leaders, 50% of investors, and 47% of the public now rank safety and security as their top AI-related concern, figures that rose from the fall 2025 wave of the same poll. The Mythos announcement was not yet a factor in that data. The direction of travel is clear: tolerance for treating AI safety as a secondary priority is narrowing.

The governance gap

No binding international framework currently governs AI systems with demonstrated offensive cyber capabilities. The AISI's assessment is exactly the kind of independent technical evaluation a governance regime would depend on, yet it arrives after the model already exists and has been selectively distributed. The institute's call for "investment now in cyber defence" is measured in tone. The underlying message is not: defenders are already behind.

Elsewhere, the industry response remains fragmented. The News reports that Google DeepMind recently hired philosopher Henry Shevlin to focus on machine consciousness and AGI readiness, a signal that even frontier labs are still assembling the conceptual vocabulary they need to reason about what their systems are doing. That groundwork matters in the long run. It operates on a different timescale than the threat Mythos represents.

Mythos closes one debate: AI-enabled offensive cyber operations are not a future-tense problem. The question now is whether existing institutions can move fast enough to address a threat that arrived ahead of the frameworks meant to contain it.

---

FAQ

What is Claude Mythos?
Anthropic's latest model, notable for its advanced cybersecurity capabilities. It can autonomously identify vulnerabilities and carry out multi-step network intrusions that would take human hackers several days to complete manually.

Why did Anthropic withhold Mythos from public release?
The company deemed the model too dangerous for general availability, citing risks to public safety, national security, and economic stability. A limited version was distributed to approximately 40 selected organizations for defensive research purposes.

What is the AISI and why does its assessment carry weight?
The Artificial Intelligence Security Institute is a UK government-backed lab tasked with evaluating frontier AI models independently. Its findings are not produced by the companies that build the models it tests.

Which organizations received access to Claude Mythos?
Anthropic confirmed access to more than 40 organizations. Named recipients include JP Morgan, Google, and Nvidia.