TL;DR — Key Takeaways
– Docker launched Cloud Sandboxes, a secure runtime designed to let AI agents run locally or in the cloud under the same controls.
– Docker also updated Kits so packaged agentic sandboxes can produce artifacts that comply with the Open Container Image specification.
– The company plans to submit Kits to the CNCF as part of an effort to establish a more standardized approach to AI agent runtimes.
Docker, Inc. this week launched a container runtime, dubbed Cloud Sandboxes, that enables artificial intelligence (AI) agents to be safely deployed in a secure runtime environment.
At the same time, Docker updated Kits, an open specification for packaging an agentic sandbox, including the agent, its tools, and the rules for what it can touch that now generates an artifact that complies with the Open Container Image (OCI) specification. Docker also revealed its intention to submit the Kits specification to the Cloud Native Computing Foundation (CNCF).
Launched at a WeAreDevelopers World Congress North America conference, Cloud Sandboxes are a more standardized implementation of a runtime based on a micro virtual machine that Docker previously developed.
Docker President Mark Cavage said that effort now makes it possible to deploy the same runtime both on an individual laptop and in a cloud service. Developers can write commands once and run them either locally or in the cloud, under the same controls, with no changes required. Docker Cloud Sandboxes are ready to go instantly, with secrets, policy, Model Context Protocol (MCP) gateways, and agent configurations already built in.
That capability will make it simpler to use any type of harness to build applications locally that also invoke remote infrastructure resources in the cloud, added Cavage. Cloud Sandboxes can scale from one to 16 fully managed virtual CPUs.
Finally, developers can more easily configure AI agents to perform tasks unattended without tying up a local machine.
Longer term, Docker is looking for more help to maintain Kits after the company devoted a significant amount of its internal resources to building it, said Cavage. The ultimate goal is to create an ecosystem of secure runtimes for AI agents that will be able to run anywhere, he added.
At present, multiple secure runtime environments for AI agents have emerged. Docker is making a case for a more standardized approach. The goal is to make it possible to securely deploy AI agents anywhere rather than, for example, having to run them in isolation on a separate physical Mac mini machine from Apple.
In the wake of a series of high-profile breaches involving AI agents, the amount of attention being paid to providing secure runtime environments to deploy them has increased. The challenge and the opportunity now is to find a way to create a runtime standard that makes it possible to safely deploy AI agents with a higher level of confidence. Otherwise, organizations will also be concerned that AI agents will, for one inexplicable reason or another, attempt to end run any guardrails that might have been put in place to prevent them from accessing sensitive data.
It’s not clear how long it might take for such a standard to be widely embraced but in the meantime, application developers building AI agents should at the very least make sure they are deploying AI agents in the most secure runtime environment possible. Otherwise, the laws of unintended consequences will all but guarantee that an AI agent will attempt to perform some task in a way that everyone affected is likely to regret.
