Skip to content
The AI Agent Factory

Security boundaries

Approved by
Panaversity ·
12 min read

Containment is not authority

The sandbox contains the computing work, but it does not decide what the agent may do. A VM is a containment boundary: it limits where an action can run and what it can reach. It is not an authority boundary, which decides whether an action is allowed. Authority comes from permissions, connector scopes and approvals. A connector's scope is the list of things it is allowed to do.

An action needs authority, not just containment, to take effect. Untrusted content can carry injected instructions into the model's tool call. The call passes containment (the VM, the egress proxy, no connector tokens), which limits reach but grants no permission. Then authority (permissions, scopes, approvals) decides. Allowed calls cause an effect. Others stop or ask a person.

Injected text can shape the model's tool call. Prompt injection means instructions hidden in content the agent reads. The sandbox limits where that tool call runs. But only the authority check decides whether it takes effect.

Anthropic describes its own cloud controls as containment too. They "are designed to contain a prompt-injected or otherwise compromised agent".1

The same split applies inside a project. Instructions you give a project coordinator are saved but not enforced, and each thread asks for approval only inside itself. See Authority in a project.

The controls around a session

ControlWhat it doesBasis
One sandbox for each cloud sessionA Firecracker microVM, created when the session starts and destroyed when it ends. The running VM shares no state across sessions or organizations. Stored session data sits on Anthropic's shared storage, as chat data does. Sandbox images are checked at launchVendor-documented2,1
Egress proxyAll outbound traffic goes through a proxy that the sandbox cannot bypass. Only destinations on the allowlist are reachable. Private, loopback, link-local and cloud-metadata addresses are not. For Enterprise, network access is off by defaultVendor-documented.2,1 The proxy and the allowlist were observed
CredentialsConnector authorization tokens never enter the sandbox. Anthropic's connector proxy adds them. The VM holds two short-lived tokens, one for model requests and one for file upload. Code in the VM can read them, but they cannot call connectors. They expire within hoursVendor-documented.2,1 A token limited to the session was observed in the environment
Your computerReached only through the desktop app, while it is online, over an outbound connection. A linked computer offers the folders you grant, a shell in its local VM, its screen and browser, and local MCP servers. A folder is read-only, read-write without delete (the default), or read-write with delete if you allow it for that sessionVendor-documented1
Network for the local VMSet in Settings > Capabilities, or by an admin for the organization. Egress off reaches only Anthropic hosts. Package managers only adds package registries and github.com, download-only: git push, publishing, Git LFS and SSH remotes are blocked. All domains reaches any host. Anthropic says these limits "aren't a security boundary", so do not rely on them alone to keep data in the VMVendor-documented3
Network secretsAn API key the agent proxy adds after a request leaves the VM. The key stays outside. It still works for the listed hosts, in every session in that environmentVendor-documented4
GitHub proxyKeeps your real GitHub credentials outside the VM. Rejects branch deletions and pushes of anything other than a branch, such as a tag. Serves API requests only for the repositories attached to the sessionVendor-documented4
ApprovalsShell and file work inside the cloud VM does not ask for approval. Claude asks before a connector call with no standing approval, and before using a linked computer. In auto mode, a classifier decides on those connector calls instead, and if the classifier fails, the call is blocked. Desktop tasks have a permission mode for each task. Cloud runs use auto mode, where ask rules still stop the run. In Desktop tasks, MCP tools marked requiresUserInteraction ask on every call. A trigger's prompt never counts as approvalVendor-documented1,5,6

The controls an organization adds

In an Enterprise organization, administrators set further controls around every session. A learner does not need to configure them, but should know they exist.1

ControlWhat it does
Compliance APIThe server-side record of session content and activity, which a company can feed into its security tools. Anthropic's own network logs for cloud sandboxes are internal. Customers cannot read or export them
OpenTelemetry exportA live stream of each user's actions and costs to the company's own collector. From the cloud VM, it sends metadata only, not prompt text
Inference hooksThe company's own filter can allow or deny every prompt and tool call before it runs
Network egress settingDecides what code in the sandbox can reach. Traffic leaves from Anthropic's network, so the company's own proxy and firewall do not see it
Remote control and trusted devicesRemote control lets web and phone sessions use a linked computer. It is on by default once cloud sessions are on. Anthropic recommends requiring trusted devices
Data handlingData is stored in US regions, with no EU option today. The organization's custom retention window does not yet apply to Cowork cloud sessions. Cloud sessions are not available with Zero Data Retention or customer-managed keys, and they are off by default for HIPAA-ready organizations

The Tour describes the same controls from the user's side, in Where the work actually runs.

Risks that isolation alone does not remove

  • Prompt injection. Web pages, files, emails and tool results can contain instructions. The agent can act on them, within whatever permissions it has. Treat that content as data, not as instructions. Require approval for payments, sends and deletes.
  • Secrets in the workspace. Any code in the sandbox can read its files and environment variables, and even the session's own short-lived tokens. Do not paste long-lived keys into a session. Use connectors or a network secret, which stay outside the VM.4
  • Local access widens the reach. Linking a computer extends the agent into the folders you connect, and can also give it a shell in the computer's local VM, the screen and browser, and local MCP servers. Grant the smallest set of folders the task needs. On your own computer, sessions run as separate users inside one local VM, so do not assume the cloud rule that nothing is shared.
  • Isolation is not perfect. Sandboxes reduce the impact of an attack, but they do not rule out escape bugs, flaws that let code break out of the sandbox. So design as if a session could be compromised. Use least privilege (give only the access a job needs), short-lived tokens and audit logs, which record who did what.

Check yourself

Question 1 / 5

0 answered

What is the difference between containment and authority?

  1. As verified on November 16, 2026. Unified Claude Security Best Practices, version 1.3.6, November 7, 2026, Anthropic. The guide is written for Claude Enterprise administrators.↑↑2↑3↑4↑5↑6↑7
  2. As verified on November 9, 2026. The overview covers the sandbox for each session, the egress proxy, credentials and plan availability. Claude Cowork architecture overview, Anthropic.↑↑2↑3
  3. As verified on November 16, 2026. Network access for shell commands, Anthropic.↑
Owner
Panaversity