UK AI Safety Institute warns GPT-5.5 poses cyber threats on par with Mythos

The UK AI Safety Institute said OpenAI’s GPT-5.5 showed strong cyberattack capabilities in controlled tests, including vulnerability discovery and attack-code execution.

In some tasks, GPT-5.5 outperformed Mitos and performed well in network intrusion simulations and CTF-style challenges.

Public models still have safety guardrails, but red-team testing found they can sometimes be bypassed.

The findings intensify concerns that advanced AI could meaningfully aid real-world hacking, increasing the need for tighter AI safety rules and stronger cyber defenses.

The UK government is responding to rising cyber threats with additional funding.