Addendum to GPT-5 system card: GPT-5-Codex (openai.com)

OpenAI published an addendum to the GPT‑5 system card introducing GPT‑5‑Codex, a variant of GPT‑5 tuned specifically for agentic coding in the Codex ecosystem. Like its codex‑1 predecessor, GPT‑5‑Codex was trained with reinforcement learning on real-world coding tasks across diverse environments to produce code that mirrors human style and PR preferences, follow instructions precisely, and iteratively run tests until they pass. The model is exposed both locally (terminal/IDE via Codex CLI and an IDE extension) and in the cloud (Codex web, GitHub integration, and the ChatGPT mobile app), enabling tightly integrated developer workflows and automated code agents. The addendum focuses on comprehensive safety measures: model‑level mitigations include specialized safety training to reduce harmful-task outputs and resistance to prompt injection, while product‑level protections add agent sandboxing and configurable network access to limit execution scope and external interactions. For the AI/ML community this signals continued specialization of large models for agentic workflows and production coding, increasing automation in development and CI pipelines, but also raising research and engineering priorities around secure execution, adversarial prompt defense, and verifiable testing. The combined RL-trained behavior and layered deploy-time controls exemplify a pragmatic approach to enabling powerful coding agents while attempting to contain misuse and runtime risk.