Shield Agent AI

All 13 questions

Question 6 of 13 · Permissions

What can the agents execute: shell commands, network calls, file writes?

Clients also ask it like this:

  • What permissions do your AI agents have on developer machines?
  • Can an agent run shell commands or make network calls without a person?
  • Describe the sandbox or limits around your coding agents.

What the client is checking

The client wants to know how much power an agent has and who set the limit. They look for a clear scope, not the words "it is safe".

Three draft answers

Pick the level that is true for you today. Every statement marked [VERIFY] must be true for your organisation before you send it. Delete a statement that is not true. Fill each {{placeholder}}.

Not yet in place

Coding agents can run the commands and edit the files that the developer who starts them can. [VERIFY] We have not set limits that are specific to agents. [VERIFY] We will define them by {{date}}. [VERIFY]

Policy only

Our policy says what agents may do: {{allowed actions}}. [VERIFY] It also says what they may not do: {{forbidden actions}}. [VERIFY] Developers set the permission mode of the agent as the policy says. [VERIFY] We do not check the settings on each machine. [VERIFY]

Enforced, with evidence

Agents run with these limits: {{shell, network and file-write limits}}. [VERIFY] Commands of the classes {{classes}} are stopped by a technical control, not by policy alone. [VERIFY] The control and its limits are described in {{document}}. [VERIFY] Every agent action is recorded and we review the records every {{interval}}. [VERIFY] We know that the control does not cover {{known gaps}}. [VERIFY]

Evidence to attach for the strongest answer

  • Exports of the agent permission settings (managed settings or the tool's admin screen).
  • The list of command classes that are stopped.
  • A redacted example of a stopped action.
  • The list of known gaps.
  • Records of the review of the activity record.

How to move up one level

From not yet in place to policy only

Decide for each class what is allowed: read files, write inside the repository, run shell commands, make network calls, touch files outside the repository. Write the decisions down. Set the permission modes of the tools to match.

From policy only to enforced, with evidence

Put the limits where one developer cannot lower them alone (managed settings, managed hooks with allowManagedHooksOnly, sandbox). A hook set by the user can be removed by the user. Stop the classes that must never happen. Record every stop. List the gaps, for example scripts that an agent writes and then runs, and commands that are built at run time.

Where KYA fits

KYA records the tool calls of the coding hosts it wraps: shell commands, file writes and MCP tool calls. kya map shows what the agents did.

The local hook is advisory and fails open. If KYA breaks or times out, the call goes ahead. It denies a fixed list of destructive commands, secret reads and exfiltration patterns (the never-list and the guard pack). It records every other call and does not block it. Hooks load when a session starts, so they apply to new sessions only.

The hook does not see everything: a script that an agent writes and then runs, commands built at run time, and built-in shells of tools that have no hook. KYA is a seatbelt, not a sandbox. An optional kya sandbox (beta) runs a command in a Firecracker VM. Its binaries are not in the npm package.

With a hosted workspace, the workspace policy decides on evaluate calls. In the hosted hook mode (preview), a call is denied when the workspace cannot answer. The hosted service fails closed.

Limits of the hosted hook: the agent runs as the user. It can edit the settings or kill the hook. Removal is detected, not prevented. Only Claude Code has onFailure: block.

Get all 13 answers in one document

Choose the level that is true for each group. You get a preview at once and the full pack by email.

Make my Answer Pack

Previous question Next question

These are drafts. Send only what is true for your organisation. This is not legal advice.