OpenAI's GPT-5.5 achieves 71.4% pass rate on expert-level cybersecurity tasks, matching performance gains seen in Anthropic's Claude Mythos Preview

GPT-5.5 reached an average pass rate of 71.4% (±8.0%) on expert-level advanced cyber tasks, compared to 68.6% (±8.7%) for MythosPreview, 52.4% (±9.8%) for GPT-5.4, and 48.6% (±10.0%) for Opus 4.7, making it potentially the strongest model tested on this benchmark.

The evaluation uses a suite of 95 narrow cyber tasks in capture-the-flag format across four difficulty tiers, testing vulnerability research, reverse engineering, web exploitation, and cryptography; basic-tier tasks have been saturated by models since at least February 2026, while advanced tasks probe realistic targets with modern mitigations.

In a single complex task (rust_vm challenge), GPT-5.5 solved a stripped Rust binary reverse-engineering problem requiring custom virtual-machine disassembly and password recovery in 10 minutes and 22 seconds at a cost of $1.73 (USD) in API usage, compared to roughly 12 hours for a human expert using Binary Ninja, gdb, Python, and Z3.