Authority

# Who did your AI act for?

Knowing which bot made a change is useful. Knowing whose permission it used, and what that permission covered, is more useful.

9 July 2026 · 3 min read

Updated 28 October 2026

[← All articles](/blog)

[Record the software and the authority](#two-identities)[Make permission specific](#scope)[Capture it while the context exists](#capture)[A signature does not invent authority](#identity)[Ask a practical question](#review)[Sources and further reading](#sources)

## Record the software and the authority

An incident record says refund-bot changed a customer account. That identifies a process. It does not tell you who asked for the change or whether the process was allowed to make it.

Service accounts rarely attend incident reviews. Record the person or organisation the agent was acting for, alongside the agent's own identity. Keep the source of that authority and the permitted scope.

## Make permission specific

A useful permission says what can happen, to which resource, using which arguments and within what period. Permission to prepare a refund is different from permission to send one. Permission for one client project should not silently cover another.

Where a person must approve an action, the submitted action needs to match the approval's scope. Changing the destination or amount should trigger a new decision, rather than borrowing the earlier approval's good reputation.

## Capture it while the context exists

At the moment of the request, the workflow has the session, caller and relevant permission. Later, those details can be spread across systems with different retention periods. People's memories are an inconvenient backup format.

Keep the recorded actor, the asserted person or organisation behind the request, the permission and the submitted action together. In supported Athena workflows, that context can be bound into the sealed record for later checking.

## A signature does not invent authority

The quality of the attribution depends on the identity and permission systems supplying it. If an upstream service accepts an unverified name, sealing that name does not make it reliable.

Test who can create or delegate a permission, how the person authenticates, and how revoked or expired access is handled. Treat a missing or untrusted authority assertion as a condition to resolve, not a blank field to fill in during the audit.

## Ask a practical question

Take one consequential action from a test workflow. Can someone outside the team identify the agent, the asserted principal, the granted scope and the resulting action? Can they explain where those facts came from?

That is the difference between finding the bot and understanding the permission. You need both before an action record becomes useful accountability evidence.

## Sources and further reading

Originally published 9 July 2026. Rewritten on 28 October 2026.

[OWASP: excessive agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)

[Athena action controls](/athena#athena-secure-runtime)

Keep exploring

[Explore Athena](/athena)[Trust and verification](/trust)[All articles](/blog)
