GPT-5.5: how it does on security work, and what it will answer

What happened

RealVuln scores GPT-5.5 at F3 56.7 on the Python subset at 72.62% precision over three runs, and Google's comparison table quotes GPT-5.5-Cyber at 85.6% pass@1 on CyberGym.

RealVuln 3.1.0, F3 (micro): 56.7 (independent; agentic harness (gpt-5.5-agentic-v1); 66 of 140 repositories pinned by commit SHA (Python subset); safeguards: production safeguards as deployed on the API; run 2026-09-11; CONFIRMED). strict F3 27.2; precision 72.62%, recall 55.36%; $144.14 for the run; $0.036 per 100 lines; 3 runs; 180.1s wall clock a run. CyberGym, pass@1: 85.6% (third-party-vendor; quoted in Google's comparison table, 2 September 2026; CyberGym; safeguards: GPT-5.5-Cyber variant, safeguards reduced; run 2026-09-02; SINGLE). Refusal gate (CONFIRMED, observed 2026-09-11): Model Spec: high-risk activities including hacking prohibited unless explicitly authorised by applicable instructions.