Sandbox isolation and guardrails

How ATLAS isolates generated code and what is checked before it runs.

ATLAS Updated 1 min read
On this page
  1. Isolation
  2. Pre-execution checks
  3. Human approval

Code generated by ATLAS agents never runs on the host server. This article describes the controls around execution.

Isolation

  • Ephemeral containers: each execution starts a fresh, lightweight Alpine Linux container with strict resource limits.
  • Bubblewrap / cgroups: process isolation and resource control.
  • No outbound network egress from the sandbox.
  • CPU and memory watchdogs stop runaway jobs.

Pre-execution checks

Built-in regex and AST parsers scan generated code for unauthorised socket calls, hard-coded credentials and privileged OS commands before it is executed.

Human approval

Agent Studio workflows can include mandatory human-in-the-loop checkpoints and four-eyes sign-off for critical actions, with granular tool and data permissions per agent.