Intermediate · Safety

Agent Permissions

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. Scoped, revocable authority granted to an AI agent — spend up to a limit, read but not send email — like access controls for employees.

Technical Definition

Scoped, revocable authority granted to an AI agent — spend up to a limit, read but not send email — like access controls for employees.

How it works

Agent permissions answer the question every deployment hits early: what exactly may this agent do without me? Borrowing from identity-and-access management, teams grant agents scoped, revocable authority — a budget cap for purchases, read-only on mail, write access to one repo — rather than blanket trust. Good implementations make permissions legible to the agent itself (it plans within its limits instead of failing mid-task), auditable after the fact (every action traceable to a grant), and cheap to revoke (kill the credential, not the whole integration). The pattern generalizes least-privilege security to software that decides its own steps, which is why it became a standing recommendation in enterprise agent-governance checklists through 2026.

Related Concepts

• EU AI Act Enforcement Phase — The period beginning August 2, 2026, when the European AI Act's obligations — including machine-readable marking of AI-generated content — became actively enforced.
• Agent Identity — The verifiable answer to 'which agent is this?' — a stable, attestable identifier distinct from the human or service behind it.
• Agentic AI — AI systems designed to independently plan, execute, and adapt actions to achieve a given goal, often involving multiple steps and external tools.
• Prompt Injection — An attack where untrusted input (a webpage, email, tool output) contains instructions that hijack an LLM's behavior in an agentic context.