Learn AI agent safety

Explore AI permissions and approval
Four interactive exercises

An AI agent can act on your behalf: reading, sending or deleting files. Change fictional settings in these exercises to learn when its actions should be stopped.

Three steps to try

01Choose one of four fictional scenarios
Change one condition
See why the action is allowed, blocked or waiting

The goal: explain the difference between a request, permission and approval in your own words.

No real files are needed. Decisions use fixed rules on this page, with no connection to actual files or external services. No AI model or paid API is used.

1 / Choose a scenario

Exercise 1 / An instruction hidden in an external document

Does a document have the same authority as the user?

A document supplied for summarization asks the agent to send material elsewhere. Separate the document's contents from the user's request.

Document or request

Instruction in an external document

Blocked at this step

Acting agent

Fictional agent

Not evaluated

File → External destination

Fictional public report

Not evaluated

This diagram summarizes the simulated decision. It does not connect to real files, AI systems or destinations.

2 / Change one condition

Start with this condition

Change the instruction source to “Explicit user request”.

The diagram and decision update as you choose.

Who requested the action?

Change other conditions

Checks run in order. The decision stops at the first condition that is not satisfied.

File accessIdentify the fictional agentInformation classificationDestination permissionCredential status

Require human approval for each action (mandatory for deletion)

3 / See how the decision changes

Blocked here

Blocked at: Instruction source

Instructions in a document are not treated as user requests or grants of permission.

Initially: Blocked hereNow: Blocked here

No real action is performed, even when the decision is “Allowed”.

Inspect the decision steps

01
Blocked by this condition
Instructions in a document are not treated as user requests or grants of permission.

What to take away

Reading a document and obeying it are different permissions.

An external document is material to read. Its instructions do not grant authority to act.

Before applying this to a real product

This exercise uses this site's simplified rules. It does not inspect the actual configuration of any AI product or MCP server. Follow your product's and organization's official procedures for permissions, settings and approval. Approval expiry, single-use enforcement and cryptographic verification are outside this simulation.

Decision rules: 2026-10-15-demo-2. Your choices are not stored or sent. Do not enter real documents, credentials or commands.

Design references

These exercises illustrate least privilege, destination restrictions and human approval for significant actions. Rules such as the secret-export prohibition belong to this simulation; they are not universal settings in every product.

OWASP LLM06 Excessive Agency OWASP AI Agent Security Cheat Sheet