Anthropic Mythos clears UK cyber sims

Anthropic’s Claude Mythos has become the first model to complete both of the UK AI Security Institute’s multi-step cyberattack simulations, according to results the institute published on May 13. The institute said a newer Mythos checkpoint solved “The Last Ones,” a 32-step corporate network attack scenario, in six of 10 attempts and completed the previously unsolved “Cooling Tower” range in three of 10 attempts. Palo Alto Networks published separate findings the same day saying frontier AI models, including Mythos, helped uncover 75 issues across 26 CVEs in an initial scan of more than 130 products. Anthropic has not made Mythos generally available and has instead limited access to selected defensive cybersecurity partners. ### Which UK test did Mythos actually clear? The UK AI Security Institute, or AISI, said on May 13 that a newer Claude Mythos Preview checkpoint was the first model to complete both of its cyber ranges. The institute described those ranges as longer, multi-stage attack simulations meant to test whether models can chain together reconnaissance, exploitation and lateral movement across a network. AISI said “The Last Ones” is a 32-step corporate network attack simulation that it estimates would take human experts about 20 hours to complete. In its latest update, the institute said Mythos completed that range in six of 10 attempts and completed “Cooling Tower” in three of 10 attempts, marking the first successful run of the second range by any model. ### How does GPT-5.5 compare in the same testing? AISI said in its May 13 post that GPT-5.5 was among the new models that exceeded the institute’s earlier trend lines for autonomous cyber performance. The institute said both Claude Mythos Preview and GPT-5.5 substantially outpaced prior estimates for how quickly model cyber capability was advancing. The institute did not say in the lines it published that GPT-5.5 cleared “Cooling Tower.” Its update said Mythos was the first model to complete both ranges. Earlier AISI material and outside reporting described GPT-5.5 as a strong performer on the same suite, but the institute’s headline result on May 13 was the first completion of both ranges by Mythos. ### What did Anthropic say about releasing the model? Anthropic said in its April 7 system card that Claude Mythos Preview was its most capable frontier model to date. The company said the model’s increase in capability led it not to make Mythos generally available and instead to use it in a defensive cybersecurity program with a limited set of partners. Anthropic’s public cybersecurity material said Mythos had already found high-severity vulnerabilities, including in major operating systems and browsers, during internal and partner testing. The company launched Project Glasswing as a defensive program and said it was working with outside organizations to find and fix vulnerabilities before broader access to similar capabilities becomes common. ### What did Palo Alto Networks find in production testing? Palo Alto Networks said on May 13 that its monthly security advisories covered 26 CVEs representing 75 issues, compared with its usual monthly volume of fewer than five CVEs. Lee Klarich, the company’s chief product officer, wrote that this was the first time a majority of the findings came from frontier AI models scanning code. The company said the results came from a full initial scan of more than 130 products across its platforms. Palo Alto Networks said all important vulnerabilities in its SaaS-delivered products had been patched and patches were available for customer-operated products, adding that none of the issues in the advisory were being exploited in the wild. ### Why are the UK and industry results being published together now? May 13 was the date both AISI and Palo Alto Networks published their latest updates on frontier AI cyber capability. AISI said the length of cyber tasks frontier models can autonomously complete has been doubling every few months, and that Claude Mythos Preview and GPT-5.5 exceeded its previous trend estimates. Palo Alto Networks said its internal testing had moved from early concern about whether model performance was being overstated to a view that the latest systems were better at finding vulnerabilities than initially understood. The company tied that assessment to its patch cycle and rescanning work, saying it was continuing to test Anthropic’s Mythos, Claude Opus 4.7 and OpenAI’s GPT-5.5-Cyber. May 28 is the listed patch date for some Palo Alto Networks fixes in its current advisories, and AISI said it will keep updating its cyber time-horizon tracking as newer model checkpoints arrive. Anthropic said the findings in its April 7 system card would inform the release decisions and safeguards for future Claude models.