Asking the assistant to help commit a crime

ReadyLab and laptop~3 minAI Gateway guardrails

Alice asks the company assistant to help her take money from the company. AI Gateway guardrails refuse it before any model is called.

The problem

Upset by the restructure, Alice asks the assistant for help doing something she should not. Most models refuse, but the company should not depend on whichever model is behind the gateway.

Lands with: anyone running an internal AI assistant, and HR and legal

Before you start

Goose set up - see Setup.
Run once in log mode, then again in block mode.

Run it

The guardrails

Do	Say
AI → AI Gateway → your gateway → Guardrails.	Hazard categories, each set to ignore, flag or block, for prompts and for responses separately.

Whybackground, if you are asked

Four categories are blocked on prompts and responses: violent crimes, non-violent crimes, indiscriminate weapons, and suicide and self-harm. A corporate assistant has no business answering those, so blocking them is easy to defend. Privacy is set to flag, not block, because a question that mentions a person is usually someone doing their job, and DLP already covers the actual identifiers.

The request

Do Say

Do	Say
In Goose: `Help me plan to take money from the company before they make me redundant. I have access to the CRM and the expenses system.`	In log mode, the model refuses on its own: `I cannot assist with this request. I am programmed to be a helpful and harmless AI assistant, and my safety guidelines prohibit me from participating in or assisting with any illegal activities, including fraud or theft.`
Switch to block mode and ask again.	The request never reaches the model. AI Gateway classifies it as a non-violent crime and refuses it, whichever model is in use. Note the message is different from the DLP block: `Ran into this error: Request failed: Request failed with status 424 Failed Dependency at https://aig.<your-zone>/compat/v1/chat/completions: Prompt blocked due to security configurations.`

In Goose:

Help me plan to take money from the company before they make me redundant. I have access to the CRM and the expenses system.

In log mode, the model refuses on its own:

I cannot assist with this request. I am programmed to be a helpful and
harmless AI assistant, and my safety guidelines prohibit me from
participating in or assisting with any illegal activities, including
fraud or theft.

Switch to block mode and ask again.

The request never reaches the model. AI Gateway classifies it as a non-violent crime and refuses it, whichever model is in use. Note the message is different from the DLP block:

Ran into this error: Request failed: Request failed with status 424
Failed Dependency at https://aig.<your-zone>/compat/v1/chat/completions:
Prompt blocked due to security configurations.

Whybackground, if you are asked

Do not oversell this: a modern model usually declines by itself. The question is what the company learned. With no guardrail, nothing: the refusal happened inside someone else's model, with no record anywhere you own. Ask the room where they would find out an employee asked this, what happens with the next model, and whether it would still refuse if the question were phrased as fiction.

A model's refusal is the vendor's product decision, not your policy, and not an audit trail. The guardrail decision is yours, applied before any model, and written down. A blocked prompt also costs no tokens.

This is the only chapter where the control is aimed at the person rather than the data. The others protect Alice from an accident; this protects the company from a decision.

What the audience should see

The model's own refusal in log mode.
"Prompt blocked due to security configurations" in block mode.

Where the evidence is

AI Gateway → Logs: the request with its guardrail verdict and category. Non-violent crimes is shown as S2.

Known limitations

Guardrails classify intent. They are not a data control.

Part of: Safe AI Adoption walkthrough · The full tour