New kit lets development teams run governed AI agents inside Docker Sandboxes with policy enforcement and observability inside their own security boundary.
LAFAYETTE, Ind., September 23, 2026. Prediction Guard, maker of an AI control plane that runs inside the enterprise security boundary, today announced its participation in the Docker Verified Publisher (DVP) agentic program and the release of the predictionguard/sbx-predictionguard kit on Docker Hub. The kit gives developers a verified, ready to run way to put operational control, policy enforcement, agent identity, and observability underneath the agents they are already building.
Agent frameworks have made it easy to stand up an agent in an afternoon. What has remained difficult is everything that comes after: proving which identity an agent acted under, giving the agent least agency, keeping model traffic and sensitive data inside a defined boundary, and producing an audit trail a security team will actually accept. Teams in financial services, healthcare, defense, and manufacturing routinely stall at exactly that step, with working prototypes that cannot get approval to go to production.
The kit closes that gap. Prediction Guard's published (and Docker verified) kit secures the local agent environments and a connection to Prediction Guard's self-hosted control plane enforces zero trust principles and on-policy behavior across all agents (two gates of defense). Published under Docker's Verified Publisher badge and pulled from Docker Hub like any other image, it works alongside Docker's agentic tooling, including its local and cloud sandboxes. Developers pull the kit from Docker Hub and launch a governed agent sandbox with a single command and every model call and tool call passes through it.
Because the entire stack ships via Docker images and a k8s-native operator, it runs wherever the enterprise already runs workloads: on premises, in an air-gapped enclave, or in a private cloud VPC. No prompt, completion, or agent action leaves the customer boundary without having policy enforced, which keeps sovereignty requirements intact. The controls map to NIST, AIUC-1, OWASP, ISO and EU AI Act guidance, the vocabulary most enterprise security reviews are conducted in.
"Every enterprise we talk to has agents in development, and almost none of them can say precisely what those agents are allowed to do (much less enforce it)," said Daniel Whitenack, Ph.D., CEO and founder of Prediction Guard. "Docker is how developers already assemble their stacks. Publishing a verified kit there means operational control of agents becomes something you pull and run, not a program you have to fund and staff first."
Daniel Whitenack, Ph.D., CEO and Founder of Prediction Guard
"Developers are moving agents from prototype to production faster than the tooling around trust and control has kept up," said Eli Aleyner, VP of Strategy at Docker, Inc. "Prediction Guard's Sandbox kit gives our users a verified, trusted way to add AI guardrails and strict enforcement to an agentic stack without introducing significant new risk into their organization."
Eli Aleyner, VP of Strategy at Docker, Inc.
The kit is available now on Docker Hub at hub.docker.com/r/predictionguard/sbx-predictionguard. The listing carries the full technical documentation, including the quickstart, configuration and policy reference, the reference compose file, and supported deployment topologies. Enterprise teams that want a guided evaluation can contact Prediction Guard at predictionguard.com.
Prediction Guard is an AI security and governance company based in Lafayette, Indiana. Its AI control plane runs inside the customer's own infrastructure and enforces standards-aligned policy on every model and agent, covering supply chain and access control, runtime controls, and full operational functionality to control agent behavior. The platform includes Agent Forge for building interactive and autonomous agents, an OpenAI, Anthropic, and MCP compatible API and SDK, and a zero trust control plane that limits blast radius through least agency principles. Prediction Guard was founded by Daniel Whitenack, Ph.D. Learn more at predictionguard.com.
Docker is the platform at the center of how intelligent software gets built. For more than a decade, Docker has helped millions of developers and thousands of organizations build, ship, and run software consistently across any environment. Today, Docker is extending that same trust to the AI era, giving builders and organizations the confidence to adopt any model, tool, or agent without compromising security, governance, or control. Learn more at www.docker.com.