The UK's Artificial Intelligence Safety Institute (AISI) just dropped a report that should make cybersecurity teams very nervous. AI's ability to autonomously hack through networks is accelerating faster than anyone predicted. Since the end of 2024, the length of cybersecurity tasks an AI can complete independently has been doubling every 4.7 months. And the latest models — Claude Mythos Preview and GPT-5.5 — have actually broken that curve.
To keep things fair, AISI limited each task to 2.5 million tokens of compute. Even with that handicap, both models achieved nearly 100% success on the hardest, 12-hour-long tasks. The report admits these models have hit the ceiling of what the current test suite can measure.
In more realistic enterprise cyber range exercises, AISI set up two attack scenarios. Claude Mythos Preview became the first model to conquer both: it passed The Last Ones 6 out of 10 times, and it was the first to crack the notoriously difficult Cooling Tower range (3 out of 10). GPT-5.5 also managed 3 out of 10 on The Last Ones.
Frontier models are no longer improving on a timescale of years — it's months now. The existing safety evaluation frameworks are being outpaced, and the window for businesses to build up their defenses is closing fast.