[Podcast] Decoding GPT-5.3-Codex

Описание: GPT-5.3-Codex System Card

https://cdn.openai.com/pdf/23eca107-a...

The provided text is a system card for GPT-5.3-Codex, a highly advanced AI agent designed for complex coding and autonomous research tasks. This model represents a significant milestone as the first release to reach a High capability threshold in the Cybersecurity domain, triggering specialized safety protocols under a Preparedness Framework. To manage inherent risks such as data destruction and malicious exploitation, OpenAI has implemented a layered safety stack including isolated sandboxes, refusal training, and real-time conversation monitors. Comprehensive evaluations across biological, chemical, and cyber domains demonstrate that while the model excels at long-horizon technical execution, it still requires rigorous human oversight and gated access for sensitive dual-use applications. The document also highlights a Trusted Access for Cyber program, which balances the need to empower security defenders while impeding potential threat actors. Ultimately, the report outlines a cautious iterative deployment strategy that prioritizes safety and risk mitigation for frontier-level AI capabilities.

#ai #largelanguagemodels #gpt #codex #openai #codingagent