How fast is autonomous AI cyber capability advancing? (www.aisi.gov.uk)

Recent estimates from the Artificial Intelligence Security Initiative (AISI) indicate that autonomous AI cyber capabilities are accelerating significantly, with the time it takes for AI models to complete cyber tasks halving to every 4.7 months as of February 2026. This marks a notable improvement from earlier estimates and is exemplified by the performance of two new models, Claude Mythos Preview and GPT-5.5, which have achieved success rates that exceed previous trends, though it's unclear whether this represents a sustainable acceleration. The implications for the AI/ML community are profound, as the rapid progression of AI in cyber capabilities translates into increasing risks and opportunities. Recent evaluations revealed that both Claude Mythos Preview and GPT-5.5 are capable of solving complex cyber tasks with unprecedented success rates. AISI’s new benchmarks, which measure task completion against human expert times, highlight that current models operate at a level of autonomy that could reshape the landscape of cybersecurity, necessitating immediate investment in stronger security measures. The potential for these models to enhance both offensive and defensive capabilities in cybersecurity underscores the urgency for organizations to adapt to this fast-evolving frontier.