Contain agents by limiting what they can do
Supervise agent capabilities via access boundaries like sandboxes and VMs, not just behavior.
Containment by enforcing access boundaries is a primary method for managing agent risk.
Rather than solely supervising what an agent does, focus on limiting what it is able to do. This is achieved through mechanisms like sandboxes, virtual machines, and egress controls. By setting hard boundaries on what an agent can reach, you reduce the potential damage from user misuse, model misbehavior, or external attacks. For instance, if credentials never enter a sandbox, they cannot be exfiltrated.
Using process sandboxes, VMs, filesystem boundaries, and egress controls to constrain where and how an agent can act.
Limit agent risk by enforcing access boundaries. This involves using tools like sandboxes and VMs to restrict what the agent can reach, preventing exfiltration of sensitive data and reducing the impact of errors or malicious actions.
- How we contain Claude across products ↗Anthropic Engineering · Anthropic Engineering · Jun 5, 2026 · paraphrased
First seen on Subvolts Jun 5, 2026. Updated Sep 11, 2026. Spotted a problem? Tell us.
- How we contain Claude across productsAnthropic · Anthropic Engineering · Jun 5, 2026
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicConstrain external content access for agentsLimit agent access to external content sources to reduce risk from untrusted data. · 10 min
- Anti-patternAgents acting beyond intended scope or contextAgents may take actions that seem reasonable to them but are not what humans intended due to lack of context. · 5 min
- WorkflowBuild an AI agent evaluation pipeline in one hourCompress weeks of AI agent testing into one hour using open-source tools and automated pipelines. · 60 min
- HeuristicAI agents need time-awareness for runtime controlAI agents require a sense of time to predict, estimate, and control their own actions and execution duration. · 5 min
- HeuristicComposable controls for agent autonomy and safetyUse composable controls that perform single jobs, enabling agents to run with autonomy at appropriate safety levels. · 10 min
- HeuristicDefine agent approval loops as product design decisionsFrame human-in-the-loop requirements for agents as product design choices, not technical limitations. · 10 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with in the sources.
- Where did this come from?
- Anthropic Engineering (Anthropic Engineering). Subvolts collected, normalized and credited it; the link above opens the original.