AISI UK: AI hacking capabilities double every 4.7 months; Claude and GPT-5.5 test metrics have "exceeded limits".

According to Beating's monitoring, a recent report from the UK's Artificial Intelligence Security Institute (AISI) indicates that AI's ability to autonomously perform cybersecurity tasks is experiencing an unexpected surge. Since the end of 2024, the length of cyber tasks that AI can independently complete has doubled every 4.7 months, with the recently released Claude Mythos Preview and GPT-5.5 even breaking through this growth curve. To control variables, AISI limited the computing power of a single task to 2.5 million tokens. However, even under these artificially weakened conditions, Claude Mythos Preview and GPT-5.5 still achieved a near 100% success rate in the most difficult tasks, which took up to 12 hours. The report acknowledges that these two models have reached the upper limit of the capabilities that existing test sets can measure. In more realistic enterprise cyber range tests, AISI set up two attack scenarios. The new version of Claude Mythos Preview has successfully breached both test ranges for the first time: in addition to achieving 6 out of 10 successes in The Last Ones test range, it became the first model to successfully breach the challenging Cooling Tower test range (3 out of 10 successes). GPT-5.5 also achieved 3 out of 10 successes in The Last Ones test range. The evolution of network attack and defense capabilities of cutting-edge models has shifted from "years" to "months". Existing security assessment systems are being rapidly compromised, and the window of opportunity for enterprises to build defenses is shrinking dramatically.