# Agent Firewall: per-operation API permissions for AI agents

> Agent Firewall lets you block reads, writes, or individual API operations per connection or workspace-wide, enforced before requests leave the sandbox.

Jun 26, 2026. Topics: Security, Integrations. https://www.generalinput.com/changelog/2026-06-26-agent-firewall

OAuth scopes and API keys are coarse. Connecting Google Sheets grants edit and delete on every spreadsheet, and an agent holding that credential inherits all of it. [Agent Firewall](/agent-firewall) moves the unit of control down to the operation: on any connection you can block whole categories (every write, every read) or individual operations, and admins can set standing rules across the workspace. Agents keep the access their job needs and lose the actions you never want them to take. Read [why scopes are the wrong unit of control](/blog/oauth-scopes-wrong-unit-of-control-for-ai-agents).

### How it works

- **Built on our integration docs.** Every documented operation carries an HTTP method, path, and read or write category, and every integration lists its domains, so rules expand into precise request matchers.
- **Enforced in the sandbox.** Outbound traffic passes through a proxy that checks each request against the rules. A blocked call gets a clearly marked 403 telling the agent which operation was denied and to ask a workspace owner.
- **Layered rules.** Workspace rules combine with each connection's own rules, so a member can't loosen a block an admin set.
- **Everywhere agents run.** The same policy applies in chat, agent workflow runs, and `geni exec` on your machine.

Find it in Settings under Agent Firewall, on each connection's page, and in the Integrations Tools tab.