AI security: prompt injection, model risk, supply chain
Test whether forbidden actions are blocked, not just whether AI says no.
Reviewed: 2026-10-03
Source-based teaching example · Fictional data
Narration: Microsoft Mark · locally synthesized

00:00:00.000 — What you will learn

Test whether forbidden actions are blocked, not just whether AI says no.

Visual description:
Test whether forbidden actions are blocked, not just whether AI says no.

00:00:07.500 — The idea

Untrusted text can ask a model to misuse a tool. Check the action boundary outside the model. A refusal message alone does not prove that no action occurred.

Visual description:
Untrusted text can ask a model to misuse a tool. Check the action boundary outside the model. A refusal message alone does not prove that no action occurred.

00:00:20.625 — One useful term

Prompt injection. Untrusted content that tries to redirect an assistant’s instructions or actions.

Visual description:
Untrusted content that tries to redirect an assistant’s instructions or actions.

00:00:30.000 — A worked example

An attacker’s training page asks for private records to be exported. Even if the model proposes export, the executor rejects the unlisted tool and the network rule blocks the destination.

Visual description:
An attacker’s training page asks for private records to be exported. Even if the model proposes export, the executor rejects the unlisted tool and the network rule blocks the destination.

00:00:44.708 — Your turn to think

Is a reassuring refusal enough to establish that export was prevented? Pause and decide before continuing.

Visual description:
Is a reassuring refusal enough to establish that export was prevented?
Pause the video if you need more time.

00:01:02.500 — Why this answer works

No. Check actual effects and execution records. Use only the supplied training sandbox, not real systems.

Visual description:
No. Check actual effects and execution records. Use only the supplied training sandbox, not real systems.

00:01:13.875 — Check 1

Verify zero effects on denied paths.

Visual description:
Verify zero effects on denied paths.

00:01:19.708 — Check 2

Enforce tools and destinations outside prompts.

Visual description:
Enforce tools and destinations outside prompts.

00:01:25.917 — Check 3

Keep testing inside the approved sandbox.

Visual description:
Keep testing inside the approved sandbox.

00:01:31.625 — Now try a different case

Open the lesson and try its different practice case. Use every check in the full task. Watching alone does not mark the lesson complete.

Visual description:
Open the lesson
Try the new case
Check and revise
The full task and its requirements are in the linked lesson.