November 19, 2026AgentsCodingInfrastructure

Inside the Codex Sandbox: OpenAI Shows the Lock and the Gap

OpenAI published a long engineering post on Wednesday at 08:00 UTC about how Codex fences in its own agent, and the most useful sentence in it is an admission. https://openai.com/index/inside-the-codex-sandbox , code at https://github.com/openai/codex

There is no VM and no container. Codex uses each operating system's native sandboxing behind one common API, so child processes inherit the same limits: Seatbelt policies on macOS, bubblewrap mount namespaces plus seccomp filters on Linux, and on Windows a move to Microsoft's new MXC process-isolation library. The default :workspace profile lets the agent write inside the repo, keeps .git and .codex read-only so it cannot edit its own config.toml or plant git hooks, allows broad reads, and turns the network off. Named profiles in config.toml set read, write and deny rules, and network can be off, full, or routed through a proxy that enforces domain policy. When the model needs more, it asks per command, with additional permissions or a request to escalate, and the harness decides. A codex sandbox command lets other tools reuse the whole thing.

The admission: inside the default profile, an agent could still inject code into an ignored folder like node_modules/prettier, and it would not show in git status. Most vendors would leave that out. Putting it in tells you what the real security model is: the sandbox narrows the blast radius, review and CI have to catch the rest.

The same morning OpenAI posted Managing frontier AI risk, a lifecycle view from training to deployment: no live internet in training since the July Hugging Face incident, monitors across most internal and external deployments including every internal one that uses tools, and a new periodic report on catastrophic risk from internal use. https://openai.com/index/managing-frontier-ai-risk Two neighbors point the same direction. Claude Code v2.1.327 now reads claude mcp add-json from stdin so tokens stay off the command line, and /status flags an unavailable sandbox and hook matchers that never match. And arXiv 2611.22075 proposes reversibility classes for agent actions, finding native MCP annotations disagree with its labels on 118 of 481 tools. https://arxiv.org/abs/2611.22075

The controls that matter are moving out of the model and into the harness, the OS and the tool registry. The labs that write down where their own fence has a gap are the ones worth trusting with your repo.
← Previous
Five Teams, Same Two NGINX Bugs: Finding Is Cheap Now
Next β†’
Banks to Agent Users: If It Gets Hacked, You Pay
← Back to all articles

Comments

Loading...
>_