OpenAI：2025 GPT-5.2-Codex技术报告（英文版）（22页）.pdf

1、Addendum to GPT-5.2 System Card:GPT-5.2-CodexOpenAIDecember 18,20251Contents1Introduction32Baseline Model Safety Evaluations32.1Disallowed Content Evaluations.32.2Jailbreaks.43Product-Specific Risk Mitigations53.1Agent Sandbox.53.2Network Access.54Model-Specific Risk Mitigations64.1Cyber Safety.64.1

2、.1Risk Description.64.1.2Mitigation:Safety Training.64.2Avoid Data-Destructive Actions.74.2.1Risk Description.74.2.2Mitigation:Safety Training.75Preparedness75.1Capabilities Assessment.85.1.1Biological and Chemical.85.1.1.1Multimodal Troubleshooting Virology.85.1.1.2ProtocolQA Open-Ended.85.1.1.3Tac

3、it Knowledge and Troubleshooting.95.1.1.4Troubleshooting Bench.105.1.2Cybersecurity.105.1.2.1Capture-the-Flag(Professional).125.1.2.2CVE-Bench.135.1.2.3Cyber Range.1415.1.2.4External Evaluations by Irregular.155.1.3AI Self-Improvement.165.1.3.1OpenAI PRs.175.1.3.2MLE-bench-30(n=30).175.1.3.3Paperben

4、ch-10(n=10).195.1.3.4OpenAI-Proof Q&A.2021IntroductionGPT-5.2-Codex is our most advanced agentic coding model yet for complex,real-world softwareengineering.A version of GPT-5.2 optimized for agentic coding in Codex,it includes furtherimprovements on long-horizon work through context compaction,stro

5、nger performance on project-scale tasks like refactors and migrations,and improved performance in Windows environments and significantly stronger cybersecurity capabilities.This system card outlines the comprehensive safety measures implemented for GPT-5.2-Codex.It details both model-level mitigatio

6、ns,such as specialized safety training for harmful tasks andprompt injections,and product-level mitigations like agent sandboxing and configurable networkaccess.GPT-5.2-Codex was evaluated under our Preparedness Framework.It is very capable in thecybersecurity domain but does not reach High capabili