AWS open-sources Strands Box: a policy-enforced sandbox for AI agents
AWS has open-sourced Strands Box, a developer-preview sandbox (Apache 2.0) for AI agents that run shell commands and call APIs without per-step human approval. It pairs OS-level containment with fine-grained policies written in Dogwood, the temporal policy language AWS released e
AWS has open-sourced Strands Box, a developer-preview sandbox (Apache 2.0) for AI agents that run shell commands and call APIs without per-step human approval. It pairs OS-level containment with fine-grained policies written in Dogwood, the temporal policy language AWS released earlier. Because policies can depend on what the agent has already done, you can express rules such as 'post to the incident channel, but no more than three times in ten minutes', enforced from outside the agent's own process. Real credentials never enter the agent's environment: they are attached at an egress gateway, so a compromised or misbehaving agent cannot leak what it never held. For teams moving agents from demos to production, this offers a concrete, inspectable control layer that is not tied to a single model or framework, and it is worth evaluating alongside IAM scoping and network egress rules.
This is a summary by our content curator. Read the original at AWS Open Source Blog: https://aws.amazon.com/blogs/opensource/introducing-strands-box-ai-agent-sandboxes-powered-by-dogwood/.