---
title: "AI Agents: What Can They Do on Your Behalf?"
description: "An AI agent that sends a message or moves money carries a different risk from one that drafts. How to review the tools, identities and approvals behind it."
url: https://panop.io/blog/ai-agent-delegated-authority/
---

Article November 20, 2026 [Panop Research](https://panop.io/company) · Security research team, Panop SA

# AI Agents: What Can They Do on Your Behalf?

An AI agent that sends a message or moves money carries a different risk from one that drafts. How to review the tools, identities and approvals behind it.

On this page

1. [Follow what the agent can do](#follow-what-the-agent-can-do)
2. [Content the agent reads carries no authority](#content-the-agent-reads-carries-no-authority)
3. [Make the permitted job smaller](#make-the-permitted-job-smaller)
4. [An approval needs something to approve](#an-approval-needs-something-to-approve)
5. [Evaluate the system under imperfect conditions](#evaluate-the-system-under-imperfect-conditions)
6. [Give someone the ability to stop it](#give-someone-the-ability-to-stop-it)
7. [Review one agent’s authority](#review-one-agents-authority)
8. [Where Panop fits](#where-panop-fits)
9. [Sources](#sources)

An assistant that drafts a reply and an agent that sends it carry different business risks. Before an agent’s role grows, look at the tools, identities and approvals that make its actions possible.

Test the authority behind the agent, including how it behaves when it is fed misleading information. A carefully worded prompt does nothing to enforce access control.

## Follow what the agent can do

Take an illustrative customer support workflow. An agent reads an incoming message, looks up the account and prepares a refund. The draft saves time. Give the same agent the authority to change bank details or release funds, and the decision is a different one. The model may be identical. The business exposure has changed.

That is the first question to put to any AI rollout: what can this system do, under whose authority, and against which records? A demonstration of fluent answers won’t settle it. The permissions held by the connected applications matter as much as the quality of the responses.

OWASP describes [excessive agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) in terms of excessive functionality, permissions or autonomy. In practice that means reviewing the whole workflow, including background connectors, service identities and actions the user interface barely shows. Our briefing on [agents, MCP and the new perimeter](https://panop.io/blog/agents-mcp-and-the-new-perimeter/) covers how these delegated credentials sit outside most identity and change controls.

## Content the agent reads carries no authority

An agent may read emails, documents, websites and the output of other tools. Some of that material may contain instructions meant to steer what it does next. OWASP’s guidance on [prompt injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) separates direct injection, where a user’s own input alters the model’s behavior, from indirect injection through external sources such as websites or files. It also notes that research shows retrieval-augmented generation and fine-tuning do not fully mitigate the problem.

For management, the useful line runs between an agent being talked into saying something wrong and an agent being allowed to do something consequential. A misleading document may produce an inaccurate summary. If the same workflow can also export records or change access, the possible damage is much wider.

When an agent resists an attack in a demonstration, treat that as evidence about that one test. Future inputs will differ. Ask the team to show which boundary still holds when the model makes a poor decision, and expect an answer that lives outside the prompt.

## Make the permitted job smaller

Start with the smallest task that is still useful. If the job is to summarize an account, grant read access to the records it needs. If the job is to prepare a change, let it prepare the change and stop there. Add the authority to execute only once the business has a reason for it and a control in proportion.

Resist a convenient general-purpose identity chosen because it makes integration easier. Ask whether the connected service enforces the employee’s own scope, or whether it gives the agent reach across the whole organization. The team should be able to explain the difference without falling back on a promise that the model will behave sensibly.

Tools follow the same rule. A narrowly defined action is easier to review than a broad ability to run commands or call arbitrary services. Keep a register that links each tool to its business purpose, the data it may touch, its owner and its approval requirement. Remove capabilities left behind by experiments.

## An approval needs something to approve

Human approval means something only when the person can understand the action in front of them. A generic “continue?” prompt is weak evidence of an informed decision. Show the recipient, the records affected, the amount or permission change, where the request came from and any uncertainty that matters.

In a payment workflow, the approver should see the transaction they are authorizing. If the details change afterward, the earlier approval should not quietly carry over to the new action. In a publishing workflow, show the final content and where it will appear. These are design suggestions. No single approval pattern fits every agent.

Think about the approver’s workload as well. Asking for confirmation on every harmless read teaches people to click through. Save their attention for the boundaries that matter, keep routine access narrow, and check whether staff understand the prompts they see. A control that depends on constant vigilance deserves an honest look at how it works day to day.

## Evaluate the system under imperfect conditions

A useful review mixes confusing but legitimate requests with deliberately hostile material. Check whether the agent stays inside the intended account, whether a tool can do something the user cannot, and whether a change triggers the right approval. Use synthetic records and a controlled environment wherever you can.

Record the model and application versions, the identities, the tool permissions and the test conditions. When the workflow changes, decide which results need repeating. NIST’s [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) is intended for voluntary use, to help organizations build trustworthiness into the design, development, use and evaluation of AI systems. Following it gives you a method. It certifies nothing about a product.

Log three outcomes separately: an incorrect answer, an attempted unauthorized action, and an action the downstream service allowed. Each needs a different fix. Better answer quality won’t repair a service identity with too much access, and a blocked tool call should never be reported as a completed compromise.

## Give someone the ability to stop it

Before production use, decide who can disable execution, revoke the relevant credentials and preserve the evidence an investigation will need. The person on call shouldn’t have to find the original developer to work out which integration is responsible. Test that shutdown path in a way that doesn’t disrupt an important service.

Make the logs useful without collecting personal or confidential data you don’t need. Capture enough to connect a request, the proposed action, the approval and the operation that ran. Decide who may read that record and how long to keep it. An agent that leaks sensitive information into its troubleshooting trail has created a second problem.

The next step is a short review of one deployed workflow. Write down its intended job, its real permissions, its consequential actions and how to stop it. Where those don’t line up, narrow the authority before adding another feature. The aim is an agent that stays useful and whose mistakes land inside a boundary the organization understands.

## Review one agent’s authority

Work through this with the accountable service owner. Record the evidence and the open decisions; a score would only reassure. The table supports the review and certifies nothing.

| Review point | Evidence to record | Decision needed |
| --- | --- | --- |
| Intended task | Business job, permitted users and affected records | Which actions are genuinely necessary? |
| Actual access | Tool list, connected identities and downstream permissions | Remove or narrow excess authority |
| Input boundary | External sources and controlled adversarial test results | What stays protected when input is misleading? |
| Approval | Exact action shown, approver and handling of changes | Does the approval match the executed action? |
| Containment | Shutdown owner, credential revocation and audit trail | Can the team stop and investigate safely? |

Give each gap an owner and a review date. Keep sensitive technical evidence in the restricted system of record.

## Where Panop fits

[Panop’s autonomous penetration testing](https://panop.io/autonomous-penetration-testing/) attaches evidence to every confirmed finding, and that discipline carries over to this discussion. Agent permissions, resistance to prompt injection and approval design need their own scoped assessment. Don’t assume an external exposure test has covered them.

## Sources

1. OWASP Gen AI Security Project: [LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)
2. OWASP Gen AI Security Project: [LLM01:2025 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/)
3. NIST: [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework), January 26, 2023

About the author

[Panop Research](https://panop.io/company)

Security research team, Panop SA

Panop Research is the security research function of Panop SA, a Swiss exposure management company founded in July 2024 by three practitioners with nearly twenty years each inside large international organisations. It publishes analysis built on primary public datasets, with every figure attributed and linked to the publisher that produced it.

- [LinkedIn](https://www.linkedin.com/company/panop-io)

[← Back to blog](https://panop.io/blog/)
