GPT‑5.2‑Codex is our maximum complicated agentic coding mannequin but for complicated, real-world instrument engineering. A model of GPT‑5.2 optimized for agentic coding in Codex, it contains additional enhancements on long-horizon paintings via context compaction, more potent efficiency on project-scale duties like refactors and migrations, and progressed efficiency in Home windows environments—and considerably more potent cybersecurity functions.
The program card outlines the great protection measures carried out for GPT‑5.2‑Codex. It main points each model-level mitigations, corresponding to specialised protection coaching for damaging duties and steered injections, and product-level mitigations like agent sandboxing and configurable community get admission to.
GPT‑5.2‑Codex used to be evaluated below our Preparedness Framework. It is extremely succesful within the cybersecurity area however does no longer succeed in Prime capacity on cybersecurity. We think present tendencies of unexpectedly expanding capacity to proceed, and for fashions to move the Prime cybersecurity threshold within the close to long run. Like different contemporary fashions, it’s being handled as Prime capacity on biology, and is being deployed with the corresponding suite of safeguards we use for different fashions within the GPT‑5 circle of relatives. It does no longer succeed in Prime capacity on AI self-improvement.
