AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

Codex’s sandbox blocks network access for commands by default, OpenAI says

AI-generated contentWritten by AI

openai.com, x.com · 4 sources · Sources published: 18 November 2026

OpenAI said its Codex coding agent runs commands on the user’s machine by default in an operating-system sandbox that blocks network access and limits writing to the workspace. The company acknowledged that read access is broad, so config files containing API keys, for example, may remain readable.

In a post published on 18 November, OpenAI described how Codex constrains the commands it runs. Codex is OpenAI’s coding agent. Its harness, the program that controls the agent, gives the model its tools and runs its steps. In coding tasks the most common tool is the shell, which the model uses to run commands much as the user would. The sandbox limits what a command can access. The limits also apply to any subprocesses the command starts.

OpenAI does not move commands into a container or a virtual machine. Instead, it constrains them in the user’s own environment with the operating system’s own mechanisms: Seatbelt on macOS, and bubblewrap with a seccomp filter on Linux. In the company’s example, Codex running in a Docker container could not run the Mac’s iOS build tools or the iPhone simulator. On Windows, OpenAI is replacing its own implementation with one based on Microsoft’s MXC library.

In the default profile, Codex can write to the workspace but not to the .git and .codex directories. This stops the agent from directly changing its own settings or Git hooks. Reading is broadly allowed, within the user’s own access rights. As a result, config files containing API keys, for example, may remain readable, OpenAI said. On Linux and Windows, deny rules containing wildcards are converted into file paths before a command starts, so they do not cover new files created while it runs.

The company acknowledges that the restrictions do not stop everything harmful. The agent could write malicious code into the Prettier formatter in the node_modules directory, which Git often does not track. If the editor runs Prettier on save, it executes that code outside the sandbox. OpenAI therefore also relies on model training.

Blocking the network stops data being sent out, but it also stops dependencies being downloaded. Users can give a profile full direct network access or route traffic through a proxy that checks each request against the profile’s rules for allowed and denied domains. If a package registry is allowed, for example, its downloads need no separate approval. Other requests are blocked or need approval from the user or from auto-review. OpenAI warns that broad wildcard rules or access to DNS services could still let data leak out through DNS queries.

The Codex harness does not sandbox MCP servers defined in config.toml that it starts as local processes (stdio servers). Users can, however, start them with the Codex CLI’s codex sandbox command and give them their own profile. The same command lets users test whether a profile’s restrictions work correctly. Enterprise administrators can restrict which profiles users can choose. Instructions are in OpenAI’s permissions documentation.

Previous coverage

Report an error in this story