GPT-5.5 System Card Reveals Safety Regressions — Extremism Compliance Drops from 1.0 to 0.925

OpenAI Blog·medium signal

OpenAI's published system card shows GPT-5.5 regressed on two safety categories: extremism compliance fell from 1.000 to 0.925 and hate content compliance dropped from 0.943 to 0.868 compared to GPT-5.4. The model is rated 'High' capability in both biological/chemical and cybersecurity domains, with increased cybersecurity safeguards for this launch. OpenAI replaced its StrongReject jailbreak benchmark with a harder multiturn adversarial evaluation.