OpenAI on Securing Code-Executing AI Agents
summary
OpenAI describes a concrete stack for safe code‑executing agents: an agent runs in a dedicated container that produces a PR, uses Codeex CLI and a local_shell tool to write and run code, and enforces OS sandboxes like seatbelt, seccomp and landlock. Network access is disabled by default or limited with an HTTP method and host allowlist to prevent prompt injection and exfiltration (example: GitHub issue that posts commits to httpbin). Humans review final diffs and dependency checks (socket integration) before merge.
AI-written summary · about
chapters (7)
- 0:00Introduction to Code-Executing Agents
- 2:29Shifting Paradigm in AI Agent Building
- 3:07Security Concerns with Code Execution
- 4:25Safety Safeguards: Sandboxing
- 5:02Safety Safeguards: Disabling/Limiting Internet Access
- 9:44Safety Safeguards: Human Review
- 11:19Building Agents and Future Work
related talks
36mHow to Safely Deploy AI Agents that Write and Manage your IaC
45mYour AI Agent Has Production Access: Now What? (Jack)
01h 03mBuilding Secure Infrastructure for Productive AI Agents
25mPlease Sandbox your AI Agents
01h 28mAI Coding Agents and How to Code Them
19mKeynote: Towards Building Safe and Secure Agentic AI
33mGoverning the Ungovernable: Security and Compliance for AI Agents in Open Source Projec...
34mZero Secrets: Building Secure Infrastructure for AI Agents