Security note · RW-2026-009

A tool may be built for one task while its token allows much more. A compromised agent inherits that wider access.

By Riftweld Labs

Security notes2 min read
Example consent screen. The task needs the first line. The token gets all four.

In this note

Where access grows.

  1. 01Consent

    A user or administrator grants access to an application.

  2. 02Delegate

    The application passes a credential or role to an agent tool.

  3. 03Execute

    The receiving service decides what that credential can read or change.

Follow the whole permission chain.

Start with a real workflow. Write down the person, the agent identity, the token's audience and scopes, and any roles the tool can take on. A narrow tool description doesn't narrow the credential behind it.

CheckQuestion
TaskWhat action did the user approve?
CredentialWhich resources and operations does the token permit?
DelegationCan this identity assume another role or call a second service?
RevocationHow is access removed when the task or session ends?

Test past the intended action.

In an approved test environment, check whether the same credential can reach a different record, action or resource. Keep the proof small and note the permission that allowed it. Don't judge real access from a scope name alone.

Controls to build.

  • Use credentials scoped to the resource and operation required by the task.
  • Keep approval and authorisation checks in the receiving service.
  • Prefer short-lived access and a tested revocation path.
  • Log the originating user, delegated identity and action so the chain can be reviewed.

Scope and limits.

This note describes a general review pattern. It assigns no CVE and doesn't claim that any specific framework is vulnerable. Check how your own identity provider, token exchange and downstream services behave.

AI agent security

Find out what your AI agents can really do.

We test the tools, tokens and services behind your agents, then show you what to narrow first.