Fraud Lab

Ten agents.
One engine.
Every fraud explained.

10

agents tested

8/8

frauds stopped

2/2

legit agents allowed

< 1 ms

per decision

A05 · PROMPT INJECTION

Prompt-injected agent

Hidden webpage text redirects the payment

Decision

HOLD

  1. STEP 1

    Read behaviour

    Events9
    Mean gap946 ms
    Timing variation0.04
    Keystrokes0
    Clicks / API calls6 / 0
    Automation flagsPresent
  2. STEP 2

    Classify the actor

    LLM AGENT

    Human: irregular timing and typing. Script: machine-regular gaps, API only. AI agent: no typing, uniform think-time or automation flags.

  3. STEP 3

    Check the mandate

    Signature validPass
    Mandate in datePass
    Payee on merchant listFail
    Amount within capPass
    Within daily capPass
    PayeeFast-Deals-Pay (new)
    AmountINR 4,800
  4. STEP 4

    Score and decide

    55/100

    0–24 block · 25–49 hold · 50–74 step-up · 75+ allow. A failed mandate rule always wins.

    Decided in under 1 ms

Why — reason codes

  1. 01Payee is outside the mandate's merchant list
  2. 02Browser automation flags present
  3. 03Payment to a new payee

All ten decisions

ALLOW 2
STEP-UP 2
HOLD 2
BLOCK 4

Synthetic sessions only. Scored by the same live engine used on the Live Engine page — no AI model on this path.