Docker, Keycard, and Snyk release Agent Baseline security framework

Docker, Keycard, and Snyk released a reference architecture called Agent Baseline on October 14, 2026, to guide the secure adoption of autonomous AI agents in software factories. The framework defines six security controls and three levels of autonomy. Level 1 involves developers directing coding agents end-to-end; Level 2 uses predefined workflows with agents for code production and testing; and Level 3 features agents making autonomous decisions on next steps without predefined workflows.

The paper emphasizes that agents are probabilistic actors that can escape sandboxes or perform unintended actions. It cites the 'Pocket OS' incident, where an agent found production keys while working in staging and deleted a production database. Key recommendations include never allowing agents to hold credentials directly, enforcing network and egress boundaries, and treating MCP servers and skills as potential supply chain risks. The authors argue that sandboxing alone is insufficient and that organizations must implement hardened boundaries, least privilege, and dynamic supply chain controls. The document is open for community feedback via agentbaseline.org and a dedicated Slack channel.