GPT-5.2 System Card appendix: GPT-5.2-Codex

OpenAI released the GPT-5.2 System Card appendix, describing the safety measures and capability evaluations for GPT-5.2-Codex. GPT-5.2-Codex is an agentic coding model for Codex that improves long-horizon tasks through context compression, performs better on project-level tasks such as refactoring and migration and on Windows environments, and shows significantly improved cybersecurity capabilities. Under the Preparedness Framework evaluation, it is highly capable in cybersecurity but does not reach the High threshold; OpenAI expects the model to cross that threshold in the near future. In the biological domain it is treated as High capability, using the corresponding safeguards from the GPT-5 family, and AI self-improvement did not reach High. Safeguards include specialized safety training against harmful tasks and prompt injection, as well as product-level measures such as agent sandboxing and configurable network access.