The UK AI Security Institute has published cyber evaluations of OpenAI’s GPT-5.5, finding that the model is among the strongest it has tested on cyber tasks and the second to complete one of its end-to-end multi-step cyber-attack simulations.
According to the institute, GPT-5.5’s results suggest that recent gains in cyber capability are not limited to a single model family. It says an earlier evaluation of Anthropic’s Claude Mythos Preview had already pointed to a step up over previous frontier systems, and GPT-5.5 appears to reinforce that broader trend across leading models.
The institute uses a suite of 95 narrow cyber tasks across four difficulty tiers to test capabilities such as reverse engineering, web exploitation, cryptography, vulnerability research, and exploitation. On expert-level tasks in its advanced suite, GPT-5.5 achieved an average pass rate of 71.4%, ahead of Mythos Preview at 68.6%, GPT-5.4 at 52.4%, and Opus 4.7 at 48.6%.
The UK AI Security Institute also tests models in cyber ranges designed to measure multi-step attack capability. In The Last Ones, a 32-step corporate network intrusion simulation modelled on an enterprise kill chain, GPT-5.5 completed the full attack chain in 2 of 10 attempts, becoming the second model to do so after Mythos Preview. In the Cooling Tower industrial control system simulation, GPT-5.5 did not complete the range, and no model has yet done so.
The institute stresses that these are controlled capability evaluations and do not necessarily reflect what is available to ordinary public users. It also notes that the current ranges do not yet include all the defensive conditions of real-world environments, such as active defenders, defensive tooling, or alert penalties.
Separately, the institute evaluated GPT-5.5’s cyber safeguards and OpenAI’s mitigations against malicious cyber use. It said expert red-teamers identified a universal jailbreak that elicited prohibited cyber content across all malicious cyber queries provided by OpenAI, including in multi-turn agentic settings. OpenAI later updated its safeguard stack, but the institute said a configuration issue prevented it from verifying the effectiveness of the final version.
The institute adds that if offensive cyber capability is emerging as a byproduct of broader gains in autonomy, reasoning, and coding, further increases in model cyber performance could follow quickly. At the same time, it notes that the same capabilities may also help defenders and points to related UK government work on cyber resilience, vulnerability management, and preparation for a possible ‘vulnerability patch wave’.
Why does it matter?
The significance of the evaluation is not only that GPT-5.5 performed strongly on cyber tasks, but that it adds to the evidence that offensive cyber capability may be improving across multiple frontier model families at roughly the same time. If those gains are being driven by broader advances in reasoning, coding, and agentic execution, then cyber risk may rise even when models are not explicitly optimised for offensive use. That makes evaluation, safeguards, and realistic testing environments increasingly important, especially as the same capabilities can also strengthen defensive work and shorten response times for cybersecurity teams.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
