AISI finds Claude Mythos Preview and GPT-5.5 outperform all tracked autonomous cyber capability trends

The UK's AI Security Institute found that Claude Mythos Preview and GPT-5.5 have substantially exceeded the doubling trend in AI autonomous cybersecurity task performance AISI had been tracking since late 2024. AISI measures this via the '80% reliability cyber time horizon,' using the task duration a human expert would need as a proxy for AI autonomy. That metric's doubling time had already accelerated from roughly eight months in November 2025 to approximately five months earlier in 2026; both models have since outperformed any trend lines AISI has measured. In its evaluations, AISI found Claude Mythos Preview showed continued improvement on capture-the-flag challenges and significant improvement on multi-step cyber-attack simulations. GPT-5.5 ranked among the strongest models AISI has tested and is the second model to solve one of AISI's multi-step cyber-attack simulations end-to-end. Separately, Claude Mythos Preview identified a flaw in HAWK, a post-quantum cryptographic algorithm for identity verification, that halved its effective security strength, and sped up attacks on a deliberately weakened version of a common internet traffic cipher several hundredfold. Anthropic has released Claude Mythos Preview only to select organizations to identify and patch security flaws. Neither vulnerability directly endangers existing systems. Whether the capability jump represents an isolated event or the start of a faster trajectory remains unclear.