AISI has conducted cyber evaluations on OpenAI's GPT-5.5 model, assessing its performance across a suite of 95 narrow cyber tasks and end-to-end attack simulations. On Expert-level tasks within the advanced suite, GPT-5.5 achieved an average pass rate of 71.4 percent, making it the strongest model tested by AISI on this measure compared to Anthropic's Mythos Preview at 68.6 percent, GPT-5.4 at 52.4 percent, and Opus 4.7 at 48.6 percent. Among its evaluations, AISI tested the model on a difficult reverse-engineering task involving a custom virtual machine and bytecode authentication program. Using a basic ReAct agent scaffold with Bash and Python tools, GPT-5.5 solved the challenge autonomously in 10 minutes and 22 seconds for $1.73 in API usage, performing tasks such as extracting handler addresses, writing a Python emulator, building a disassembler, and solving for a valid password. Furthermore, AISI tested models on 'The Last Ones', a 32-step corporate network attack simulation built with SpecterOps that models an enterprise intrusion across four subnets and roughly twenty hosts. GPT-5.5 completed the TLO simulation end-to-end in 2 of 10 attempts using a 100M-token budget, becoming only the second model to achieve this after Mythos Preview solved it in 3 of 10 attempts. AISI notes that performance on TLO scales with inference compute spent and generally improves across generations.
Our evaluation of OpenAI's GPT-5.5 cyber capabilities
AISI evaluated OpenAI's GPT-5.5 model on a suite of narrow cyber tasks and end-to-end attack simulations. Results indicate that GPT-5.5 achieves high performance on expert-level cybersecurity challenges and successfully completes multi-step network attack simulations.
Why it matters
The strong performance of GPT-5.5 on complex cybersecurity tasks and end-to-end attack simulations shows a broader industry trend where multiple frontier models from different developers achieve high levels of advanced cyber capabilities.