AI Agent Permissions

OAuth scopes say what an agent can call. SMERC asks what it should do now.

An AI agent may have a valid identity, approved credentials, and access to a tool while still proposing an action with poor rollback options, broad impact, weak evidence, or high external side effects.

SMERC adds a runtime permission decision that can allow, throttle, freeze, deny, or escalate the proposed action before it executes.

Permission Gap

Access control is necessary, but incomplete

IAM, OAuth, role-based access, policy engines, and agent identity answer important questions: who is acting, what tool can be called, and whether the request matches known policy. They do not always answer whether this specific action is recoverable under current conditions.

SMERC evaluates action metadata, evidence source, reversibility, containment, rollback latency, impact scope, anomaly pressure, and authorization confidence before execution.

Example Actions

Actions that need runtime permission checks

  • Deploy infrastructure from an AI-assisted pull request.
  • Delete or migrate production data.
  • Change IAM, secrets, tokens, or service-account authority.
  • Submit a write action through an MCP server or workflow tool.
  • Send external customer, financial, compliance, or operational updates.

First Proof Path

Start with GitHub Actions shadow mode

The current public pilot path evaluates AI-assisted code, deployment, and infrastructure changes in observe mode. Teams can compare SMERC posture against reviewer judgment before any enforcement claim is made.

Open GitHub Actions pilot Open 5-minute proof Review console