OpenAI has released detailed technical documentation on the sandboxing architecture for its agentic SI tool, Codex. The update clarifies how the system restricts command execution within a user’s existing environment, aiming to balance the utility of SI agents with security constraints.

What Happened

The company describes a "defense in depth" approach where the Codex harness implements specific protections alongside model alignment. The core of this security model is the Codex sandbox, which limits what programs can access while running. Unlike virtual machines or containers that isolate computation in separate environments, Codex uses each platform’s native sandboxing facilities to constrain commands in the user’s local environment. This design allows SI agents to interact with local tools, such as iOS build chains on macOS, which would otherwise be inaccessible in a standard containerized workflow.

The sandbox operates through permission profiles that define filesystem and network restrictions. For example, the built-in :workspace profile allows writes within the project directory but keeps configuration directories like .git read-only. It also disables network access by default to prevent data exfiltration, though users can configure proxies to allow specific domains. The documentation notes that while these restrictions prevent many risky behaviors, they do not eliminate all threats; for instance, an agent could still inject malicious code into locally installed packages that are outside the direct sandbox scope.

Why It Matters

For developers and enterprises integrating SI agents into their workflows, understanding the boundaries of local execution is critical for security and compliance. By exposing the sandbox mechanism through a common API across macOS, Linux, and Windows, OpenAI allows organizations to tailor permissions to specific needs. Enterprise administrators can use managed requirements to constrain which permission profiles users can select, ensuring that SI tools adhere to corporate security policies. The ability to reuse the sandbox command in other tools also suggests a broader ecosystem where SI agents can be integrated into custom development environments with granular control.

The technical implementation varies by operating system. On macOS, Codex utilizes Apple’s Seatbelt mechanism, despite it being marked as deprecated in system manuals. On Linux, it combines bubblewrap for filesystem isolation with seccomp-bpf for system call filtering. The network proxy component, which enforces domain rules, is shared across platforms but integrates with OS-specific networking stacks. This layered approach ensures that SI agents operate within defined limits, reducing the risk of unintended file modifications or data leaks during automated tasks.

The Bottom Line

OpenAI’s detailed breakdown of the Codex sandbox highlights the industry’s focus on securing agentic SI tools by constraining their local environment access. The reliance on OS-native sandboxing features allows for greater flexibility than containerized approaches, enabling SI agents to work with local development tools while maintaining strict filesystem and network controls.