How We Contain Claude Across Products
- Source
- anthropic.com
- Author
- Anthropic Engineering
- Date
Context
An 's potential damage depends on what it can reach as well as how often it makes a mistake. Anthropic's containment account explains how sandboxes, virtual machines, filesystem boundaries, and network egress controls restrict the resources available to Claude. These controls complement auto mode's approval checks by limiting the consequences of an action that gets through.
For claude.ai code execution, the article describes isolated server-side containers with a per-session filesystem and no access to the user's local filesystem. That arrangement keeps the potential damage small, but the article notes the ceiling on what Claude can do is correspondingly low, with no persistent workspace.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- human-in-the-loop — Requiring a person's approval at specific points in an automated process, chosen so the irreversible steps are the ones a human sees.
“Today that level of access is routine, and Anthropic developers are more productive for it.”
“Our telemetry showed users approved roughly 93% of permission prompts.”
“The more approvals a user sees, the less attention they pay to each, becoming over time much less diligent in their supervision.”
“More capable models make fewer mistakes, but they’re also better at finding unexpected paths to a goal, often by routing around restrictions nobody thought to write down.”
“Rather than supervising what the agent does, we supervise what it’s able to do by enforcing access boundaries through, for example, sandboxes, virtual machines, and egress controls.”
Checking sign-in…
Loading comments…