WASHINGTON— Anthropic’s Claude Mythos Preview is showing a new level of offensive cyber capability that credit unions and other financial institutions should view as an immediate, not theoretical, risk, with new testing from the U.K.’s AI Security Institute finding the model could autonomously complete a full corporate network intrusion from reconnaissance to total takeover in a controlled environment, according to CareersInfoSecurity.
CareersInfoSecurity reported the U.K. institute said Mythos Preview marks “a step up over previous frontier models” in cyber performance, continuing a rapid acceleration CUToday.info has been tracking in recent weeks around the model’s potential to compress what once took skilled human attackers days into hours. In one key test, the institute said Mythos successfully carried out a multi-stage attack sequence that it estimates would take a skilled human about 20 hours.
At the same time, CareersInfoSecurity noted the results came with important caveats. The AI Security Institute said the simulated enterprise environments lacked many features common in real-world networks, including active defenders, security tooling and penalties for triggering alerts, meaning it cannot conclude Mythos would perform as well against a hardened, closely monitored system. The model also failed to complete a separate industrial control systems simulation, appearing to stall in conventional IT portions of the test rather than the industrial components themselves.
Still, the trendline remains troubling. CareersInfoSecurity reported Mythos achieved a 73% success rate on expert-level capture-the-flag challenges—tasks no model had successfully completed before this month—and in the institute’s more complex “The Last Ones” corporate intrusion test, it completed the full sequence in three of 10 attempts while averaging nearly two-thirds of the required steps across all runs. The next-closest model completed only about half of the steps, suggesting Mythos is opening a wider gap in offensive cyber capability.
CareersInfoSecurity said the institute explicitly warned that “small, weakly defended and vulnerable enterprise systems” where an attacker has already gained access should now be considered within the reach of autonomous AI attack workflows—a warning that lands especially hard for smaller institutions and vendors with thinner security staffing.
CareersInfoSecurity further reported the institute is now rethinking how it tests frontier AI models because legacy evaluation environments may already be too easy to meaningfully measure the most advanced systems. The agency said future testing will add active monitoring, endpoint detection and live incident response, while also warning that Mythos’ performance likely improves when given more computing resources and longer run times.