Daily Tokens
November 18, 2026

AI daily use case · November 18, 2026

Running coding agents inside external sandboxes

People are letting coding agents execute generated code in a dedicated sandbox while the agent itself runs outside it, so a mistake or prompt injection cannot damage the host machine. The approach lets people try local models for agentic coding with less risk. OpenAI's Codex sandbox follows the same principle, enforcing filesystem and network limits across macOS, Linux and Windows. It has caught on because agents now run long, multi-step coding jobs, like the curl port and bugfix work, that need guardrails to be trusted.

Go deeper

More from AI on Wednesday, November 18