Securing Persistent Agents Like OpenClaw at OpenAI

OpenAI blog post on running hundreds of always-on OpenClaw agents safely: trusted control plane, sandbox, auto-review, agent identity and logs.

Source Created Updated External site Also in German: Securing Persistent Agents Like OpenClaw at OpenAI

Reviewed by adrian on

A post by Kevin S Lin on the OpenAI developer blog, published on 12 November 2026. It describes how OpenAI runs persistent agents across the company, including hundreds of OpenClaw agents; more than 500 agents now work in the background there. It says that OpenClaw launched almost a year earlier.

Content

  • The risk: persistent agents run all the time, act without being prompted and can write any code to reach a goal. The main danger is the agent doing something misaligned, not only attackers.
  • Threat model: the deterministic infrastructure (network, identity, sandbox, the gateway that configures agents) is trusted; the agent, its environment and the code it writes are not.
  • Five measures: separate the gateway from the agent harness, set hard sandbox boundaries and let a second model review actions (auto-review). The sandbox covers the network allowlist, the workspace, the process and a credential broker for secrets. The agent also gets an identity of its own, and every configuration change and every agent turn is logged.
  • Configuration: two files of the Codex harness, an admin file that sets limits (requirements.toml) and the agent's configuration within them (config.toml).
  • Outlook: OpenClaw Enterprise (OCE), an open-source, vendor-neutral control plane announced at DevDay, in pilot with version 1.0 planned by the end of the year.

Not covered: memory, prompt injection, vetting skills and plugins, updates and incident response.

Securing a Persistent AI Agent explains the post in my own words and what I take from it.

Based on

  • Explaining OpenAI's post on securing persistent agents (raw/2026/2026-11-15/explaining-openai-s-post-on-securing-persistent-agents.md) [chat]

More on this

Cited in

  • Securing a Persistent AI Agent How to let an AI agent work on its own without handing it the keys: OpenAI's five safeguards explained, and how I apply them to my own agents.

By

  • Kevin S Lin Works at OpenAI and builds tools for humans and agents. Founder of the open-source note-taking tool Dendron.
  • OpenAI American AI company in San Francisco, a public benefit corporation, that develops ChatGPT, Codex and the GPT models.

Show in the graph view