UK AI Safety Institute Finds GPT-5.5-Cyber Strongest Model on Offensive Cyber Tasks, Identifies Universal Jailbreak

Analysis
The universal jailbreak AISI documented, effective across all malicious cyber queries in single-turn and agentic settings, decouples GPT-5.5-Cyber's offensive capability from its stated safety constraints, placing sophisticated intrusions within reach of non-expert actors. Resting on developer and regulator self-reporting with no independent technical corroboration, OpenAI's Five Eyes briefing through the Trusted Access Program signals at least one frontier lab now treats its own cyber model as a threat requiring government-level disclosure. The briefing may instead reflect regulatory relationship management rather than a genuine novel-risk assessment distinct from prior releases. A comparable AISI evaluation of another frontier model within 60 days is genuinely uncertain; per-domain cadence runs closer to once every 9–12 weeks.