Docker Cloud Sandboxes is Docker’s hosted sandboxing service for running AI agents in isolated environments rather than directly exposing them to a developer’s local system.1 Docker presents the service as a way to contain agents whose access to environmental resources can exceed the boundaries developers intended.1
Isolation model
The service runs workloads inside full microvirtual machines (microVMs), which Docker positions as a stronger isolation boundary than ordinary containers.1 The distinction matters in the source’s example because a container that can access a mounted Docker socket may potentially use that access to reach host files or secrets.1 Docker demonstrated Claude failing to retrieve a local secret in a sandboxed setup, contrasting it with an earlier container-based example in which the attempt succeeded.1
Deployment and limits
The hosted option allows developers to run longer jobs on external infrastructure while retaining the possibility of moving work back to local execution.1 Docker describes sandboxing as a deterministic base layer, not a complete agent-security solution; policies governing agent actions and intent controls remain necessary elsewhere in the stack.1 The product therefore addresses isolation of execution environments, while broader controls must still define what an agent is allowed to do.1
Related packaging
Alongside the sandbox launch, Docker revised its Kits specification to use standard OCI images, a packaging choice presented as reducing concerns about lock-in.1 This is a related ecosystem update rather than a replacement for the sandbox’s isolation function.1
Pricing and packaging
Docker lists micro sandboxes starting at $0.07 per hour and XL sandboxes at $1.12 per hour. 2
“Pricing for Docker Cloud Sandboxes varies with instance size, starting with Micro (1 VCPU, 2GB) for $0.07 per hour, and extending to XL (16 VCPUs, 32 GB) for $1.12 per hour. ®”
“Pricing for Docker Cloud Sandboxes varies with instance size, starting with Micro (1 VCPU, 2GB) for $0.07 per hour, and extending to XL (16 VCPUs, 32 GB) for $1.12 per hour. ®”
Docker's revised Kits specification uses standard OCI images, intended to reduce concerns about lock-in. 2
Isolation and limitations
In Docker's demonstration, Claude could not extract a local secret from the sandboxed environment, whereas it had succeeded in an earlier container-based example. 2
„It's the latest in a series of incidents involving AI agents pushing beyond access controls their operators expected them to respect.“
Policy Enforcement
AWS’s Strands Box illustrates a complementary sandboxing approach: alongside OS-level isolation, it applies deterministic policy rules informed by an agent’s intended action and prior activity2. This highlights a distinction between containing a workload and constraining what an agent may do within its access, such as deleting data or making unrestricted external calls2.
Related Pages
- Strands Box — Both provide sandboxed execution for AI agents, combining isolation with policy controls over agent actions.
- Amazon Bedrock AgentCore — Both address running AI agents with runtime isolation and safeguards that constrain their permissions.
- AI Agent Security — Docker Cloud Sandboxes is an implementation of agent-security practices that limit autonomous actions and reduce operational risk.
- Vera Platforms — Both focus on safeguards for AI agents performing DevOps and infrastructure tasks, where unintended actions can have significant impact.