UK AISI: AI hacker capabilities double every 4.7 months, Claude and GPT-5.5 test scores have already "broken the scale"

Crypto World News reports that the UK Artificial Intelligence Safety Institute (AISI) latest report shows that AI’s ability to independently carry out cybersecurity tasks is seeing an unexpectedly explosive surge; since the end of 2024, the length of network tasks that AI can complete on its own has doubled every 4.7 months. In the recently released Claude Mythos Preview and GPT-5.5, the success rate in the highest-difficulty tasks lasting 12 hours is close to 100%. In enterprise network range testing, Claude Mythos Preview achieved successful breaches of two ranges for the first time: in the The Last Ones range, it succeeded 6 times out of 10, and in the high-difficulty Cooling Tower range, it succeeded 3 times out of 10. GPT-5.5 also achieved 3 successes out of 10 attempts in the The Last Ones range. The report notes that the evolution of cyber offense-and-defense capabilities of leading-edge models has shifted from “measured in years” to “measured in months,” and the existing security assessment system is being rapidly breached, with the window for enterprises to build defenses narrowing sharply.