GPT-5.3-Codex System Card [pdf] (cdn.openai.com)

OpenAI has unveiled the GPT-5.3-Codex, a robust evolution in AI coding models, which combines advanced coding capabilities with enhanced reasoning and professional knowledge. This model is particularly noteworthy for its potential in handling long-running tasks that require complex research and tool usage, allowing for a more collaborative interaction akin to working with a human colleague. OpenAI is cautiously categorizing this model as "High capability" in the cybersecurity domain, prompting the activation of a comprehensive suite of safeguards designed to mitigate risks and enhance security. Significantly, GPT-5.3-Codex introduces enhanced safety mechanisms, including an "agent sandbox" feature that isolates the model in secure environments to prevent unauthorized access to sensitive data. This is paired with a model-specific focus on avoiding destructive actions, where the model has been trained to recognize and mitigate the risks of executing commands that could result in data loss or corruption. The overall advancements aim to provide developers with powerful tools while emphasizing the importance of safety and user control, highlighting OpenAI's commitment to responsible AI deployment amidst growing concerns around AI capabilities and security threats.