What changed

AWS documented a preview Linux sandbox in which DevOps Agent can write and run code during investigations. Operators opt in per Agent Space, control outbound networking with an allowlist, preinstall pip and npm packages, and expose filesystem skills that may include executable code.

Why it matters

Operational agents gain a real execution environment instead of reasoning only over retrieved telemetry. That improves diagnosis while making network egress, package provenance, skill code and the sandbox-to-account permission boundary part of the incident-response threat model.

What to do

Create a disposable Agent Space with synthetic telemetry, no production credentials and an empty egress allowlist. Add one approved package and one test skill, then verify filesystem isolation, denied network access, read-only AWS calls, logs, teardown and behavior when skill code requests a forbidden action.

Caveat

The preview scope and controls are AWS documentation, not an independent containment assessment. In the July release the sandbox was limited to investigations; packages and skill code expand the supply-chain surface, and allowlists do not by themselves prevent data leakage to approved destinations.