How fast is autonomous AI cyber capability advancing?

AISI’s latest analysis indicates that the length of cyber tasks frontier AI models can autonomously complete has been doubling every 4.7 months since late 2024, a faster rate than the 8‑month doubling noted in November 2025. The agency’s February 2026 estimate shows that models’ 80‑percent reliability time horizons have been accelerating, with new models Claude Mythos Preview and GPT‑5.5 substantially exceeding previous trends. However, AISI cautions that it is unclear whether these models represent an isolated break or a new, faster trend. The organization measures progress using a narrow cyber suite and time‑horizon benchmarks, constraining each task to 2.5 million tokens to keep results comparable. This token limit understates what models can achieve, as higher token budgets improve performance. AISI also notes that longer tasks rely on only six examples, and human baselines are imperfect, adding uncertainty to the doubling‑time estimates. Despite these caveats, the agency observes consistent rapid growth across models and methods. In addition to the narrow suite, AISI’s cyber ranges test models’ ability to conduct attacks on small, undefended enterprise networks. The newer Mythos Preview checkpoint completed both ranges, solving “The Last Ones” in 6 of 10 attempts and “Cooling Tower” in 3 of 10, while GPT‑5.5 solved “The Last Ones” in 3 of 10. These results illustrate that newer checkpoints can yield significant capability jumps without new model releases. AISI concludes that autonomous cyber and software capability is advancing quickly, but the pace of future progress and its translation to defended real‑world systems remain uncertain. The findings underscore the need for governments and businesses to prepare for tangible risks posed by rapidly improving AI cyber tools.