AISI says GPT-5.5 has reached the same cyber-eval band as Claude Mythos Preview. On the expert-level tasks, GPT-5.5 averaged 71.4% pass rate, slightly ahead of Mythos Preview’s 68.6%, and AISI says it may be the strongest model it has tested.
That is a useful signal, but not a license to overread it. AISI is describing controlled tasks and simulated ranges, not a live breach of a defended target. The institute itself says those ranges lack active defenders and alert penalties, so the result is about capability, not proof of field performance.
The other detail worth keeping in view is the safeguard finding. AISI says its red-teamers found a universal jailbreak in the cyber safeguard stack. So the story is not just that the model got better; it is that capability and containment are still moving at different speeds.
Takeaways
- GPT-5.5 now sits in AISI’s top cyber-eval band alongside Claude Mythos Preview.
- The result is controlled capability evidence, not live-breach evidence.
- Safeguards remain a separate, still-open problem.
Source basis
- AISI: GPT-5.5 cyber capabilities — Official evaluation describing GPT-5.5’s cyber-task and cyber-range results, plus safeguard red-teaming findings.
- AISI: Claude Mythos Preview cyber capabilities — Earlier official evaluation used as the comparison class for the new GPT-5.5 finding.
Why operators should care
AISI says GPT-5.5 reaches Claude Mythos Preview-level cyber performance, but the result is still a controlled capability signal rather than proof of live-breach performance.
What changed
AISI says GPT-5.5 reaches Claude Mythos Preview-level cyber performance, but the result is still a controlled capability signal rather than proof of live-breach performance.
It gives a crisp read on where frontier-model cyber capability is moving, and where the safeguard problem is still lagging.
It gives a crisp read on where frontier-model cyber capability is moving, and where the safeguard problem is still lagging.