OpenAI has launched GPT-5.5, a model aimed at agentic work and complex task completion. Alongside it comes GPT-5.5-Cyber, a specialized variant capable of penetration testing, bug finding, exploit development, and malware analysis — and OpenAI is rolling it out only to a select group of trusted cyber defenders.
The UK’s AI Security Institute tested it independently and found GPT-5.5-Cyber is only the second model ever to complete one of their multi-step cyber-attack simulations end to end. On expert cyber tasks, it scores 71.4% — roughly on par with Anthropic’s Claude Mythos Preview at 68.6%. One benchmark result that’s hard to ignore: GPT-5.5-Cyber solved a challenge in 10 minutes and 22 seconds, at a cost of $1.73, that took a human expert 12 hours.
The restricted rollout is awkward given that Sam Altman recently criticized Anthropic for doing essentially the same thing with Mythos, calling that kind of exclusivity “fear-based marketing.” OpenAI is now doing a version of what it mocked, though the company would presumably argue the intent — getting the model to defenders before attackers — justifies the approach.
Whether that framing holds up is a fair question. But the underlying capability numbers are hard to dismiss.

Collections

OpenAI has released GPT-5.5 and a restricted cybersecurity variant called GPT-5.5-Cyber, capable of penetration testing, exploit development, and malware analysis. The UK's AI Security Institute found it is only the second model to complete a multi-step cyber-attack simulation end to end, scoring 71.4% on expert cyber tasks — comparable to Anthropic's Claude Mythos Preview. Notably, it solved a challenge in under 11 minutes for $1.73 that took a human expert 12 hours. The restricted rollout to vetted defenders draws irony given Sam Altman's recent criticism of Anthropic for similar exclusivity with Mythos.

OpenAI announces GPT-5.5