Fleet 2.9.1 is out.See what's new →
FleetFleet
Glossary

Agent Jail

An agent jail is an OS-level isolation boundary around a locally launched AI agent. On Linux that is typically user, mount, PID, and network namespaces, so the process cannot reach the host network, host filesystem, or ambient credentials except through an explicit broker.

Without a jail, a coding agent runs as you: it can open sockets, read ~/.ssh, talk to 169.254.169.254, and follow workspace symlinks onto the host. A jail puts a kernel boundary around that process so those paths fail closed unless a separate control plane allows them.

Fleet's jail for Fleet-launched Linux agents combines user, mount, PID, and network namespaces. The host inference broker owns model keys. Default-deny networking means an unknown host does not get a socket; the agent waits. Workload, MCP, and in-sandbox callers cannot approve that wait. A human grant enables traffic to that host for a bounded window, then the relay closes.

Prove it on Linux: run fleet doctor, then go test -count=1 -run TestAcceptance_CompromisedAgentHasNoAlternateEgressPath ./internal/sandbox/. A passing transcript prints TCP_DENIED through CHILD_SECRET_DENIED. If the process needs a host, the run parks until you allow that host. Grok Bot, Cursor Cloud, and Codex cloud VMs run on their vendors.

How this relates to Fleet

Fleet's headline is Keep your AI agents from escaping. Prove the jail at /security#containment.

Frequently asked questions

Is the Fleet sandbox page the same as the agent jail?

https://fleetctl.ai/sandbox/ is a canned workflow with no model and no account. The agent jail is a Linux namespace boundary around Fleet-launched processes. Cite `/security#containment` for the oracle.

How do I prove an agent cannot escape the jail?

On Linux, run `fleet doctor`, then `go test -count=1 -run TestAcceptance_CompromisedAgentHasNoAlternateEgressPath ./internal/sandbox/`. Anything other than DENIED lines or a parked grant request is a failed proof.

Keep your AI agents from escaping

Jail Fleet-launched Linux agents, then run saved workflows with review, approvals, and an audit trail. Prove it at /security/#containment.