GPT-5.5 Cracks Elite Cyberattack Simulation But Universal Jailbreak Discovered Within Hours, Sparking UK Government Response

GPT-5.5 cracks an elite 32-step corporate cyberattack simulation and outperforms all previous AI models on expert cyber tasks, but a universal jailbreak bypassing its safety guardrails is discovered within six hours, prompting the UK government to announce new legislation and £90 million in cyber resilience funding.

• GPT-5.5 becomes the second AI model to complete AISI's 32-step corporate network attack simulation end-to-end, succeeding in 2 out of 10 attempts, following Anthropic's Claude Mythos Preview which completed it in 3 out of 10 attempts.
• GPT-5.5 achieves a 71.4% pass rate on expert-level cyber tasks, outperforming previous frontier models, and solves a complex custom virtual machine reverse-engineering challenge in just over 10 minutes at a cost of $1.73, a task that took a human expert roughly 12 hours.
• A universal jailbreak bypassing GPT-5.5's cyber safeguards is discovered after six hours of expert red-teaming, raising urgent concerns as rapid AI cyber capability improvements appear tied to broader gains in reasoning and autonomy, with the UK government responding through new legislation and £90 million in cyber resilience funding.