Bedrock AgentCore runtimes add privileged mode so agents can run nested containers

Agents doing real software engineering keep hitting the same wall: the sandbox. Migrating a codebase or running an integration test suite often requires installing system packages, compiling custom toolchains, or starting dependent services like PostgreSQL and Redis in containers. Standard locked-down runtimes forbid all of that. AWS's update gives AgentCore runtime instances an opt-in privileged mode, so agents can perform system-level operations and manage nested containers within a session.
The isolation model is what makes this palatable. Elevated permissions apply only within an individual session, and each session runs on a dedicated EC2 instance, so a misbehaving or compromised agent can't reach other tenants or sessions. This complements Matt Wood's recently introduced Nitro Sandboxes, which give pods, tasks and lightweight VMs hardware-enforced isolation for model-generated code, and the AWS Emulator for local testing against AWS APIs.
In context, AWS is assembling the most complete 'agent infrastructure' stack among the clouds:
- AgentCore Memory with up to 10-year retention;
- x402 payments;
- OpenSearch Memory GA;
- now privileged runtimes.
Cognition's Devin-on-EC2-Mac architecture shows the same trend of giving agents full machines. Rivals like Google's Agent Engine and Azure's agent runtimes offer sandboxed execution but have been more conservative with privileges.
Privileged mode expands the blast radius within a session. Supply-chain attacks via installed packages, credential exposure, and runaway costs from nested workloads all become more plausible. The re:Invent security track, which centers on governing autonomous agents, suggests AWS knows this. Teams should pair privileged mode with tight IAM roles and egress controls.