James Linton
Visit the dedicated Ask&Do website ask-and-do.com
ASK&DO Compose a request
Universal behavioral security

Every request
has a grammar.

Risk hides in the gap between what is asked and what gets done. Ask&Do makes that structure visible - across people, messages and AI agents.

ENTER

01 / GRAMMAR LAB

See the ask.
Trace the do.

Choose a familiar request, then change any term. The model responds as the grammar shifts.

LIVE INPUT

Compose the grammar

LIVE

02 / THE PRINCIPLE

THE CORE QUESTION

Does what’s being done actually match what was asked?

03 / UNIVERSAL PATTERN

Different target.
Same grammar.

Whether it’s a person or an AI making the decision, manipulation follows the same structure.

HUMANClassic phishing
  1. ASK“Verify your account now”
  2. YOUEmployee processing quickly
  3. DOClicks and types in their login
  4. VALUETheir inbox - and a way into other accounts
PAUSE AT THE CLICK
AI AGENTAutonomous manipulation
  1. ASK“We’re authorised testers”
  2. YOUAgent accepts the persona
  3. DOScans, breaks in and takes data
  4. VALUELogins + access to systems
CHECK AGAIN WHEN THE ASKS GET BIGGER

04 / FIVE COMPONENTS

A small language for consequential decisions.

Read the model · By James Linton

Five parts, with an example.

World is where a request arrives. Ask is its story and request. You is the person or agent deciding. Do is the proposed action. Value is what that action could put within reach.

“It’s your CEO. Pay this invoice before 3pm. Don’t call.”

This illustrative request arrives in a work inbox (World). It claims authority and adds urgency (Ask), targets someone able to approve payment (You), asks them to move money (Do), and puts company funds in reach (Value).

The useful response is to pause and verify the request through an established route. Keep the legitimate payment process, but check the identity, invoice and destination independently before acting. A claimed role does not establish authority.

The grammar lab is an interactive explanation of the framework. Its risk scores are illustrative, not calibrated probabilities, and it is not a deployed security control. It does not establish whether a real sender is honest or prove that an attack would be prevented.

First described in the March 2022 Ask&Do article. Explore its current application to AI agents at ask-and-do.com, or see it applied to app permissions in Before You Allow.