Tools · Adversarial assessment

PROBE

PROBE is adversarial assessment research for AI systems, with a public workbench you can step through.

PrototypeSynthetic demo
Checked 

Research prototype. The demo uses fictional scenarios and fixed responses.

Open the synthetic workbench (opens in a new tab)Ask about PROBE

What it is

PROBE is adversarial assessment research for AI systems, with a public workbench you can step through.

The workbench walks through attack scenarios one step at a time: the input, the system’s response and the finding. Drafts stay in the page’s memory and are left out of exports.

Why it exists

Built to find out how your system behaves before someone else does.

The prepared assessment workflow has passed a code review. The report it is being built toward maps findings to the OWASP GenAI and Agentic Top 10.

Intended flow, not results

How it works

What the workbench steps through, and the report it is being built toward.

  1. Step 1

    Scenario

    A fictional attack scenario is selected.

  2. Step 2

    Input

    The input sent to the system under test.

  3. Step 3

    Response

    The system’s response. In the demo it is fixed.

  4. Step 4

    Finding

    The finding drawn from the response, with its evidence.

  5. Step 5

    Report

    Intended: a scored report designed for an outside reader to check. Not built yet.

In the public workbench, scenarios are fictional and responses are fixed.

Real pixels, dated

Inside PROBE

The PROBE evidence workbench showing a synthetic case, a scripted response and the evidence behind a verdict.
PROBE synthetic workbench · captured 7 Oct 2026. A browser demo that replays fixed, fictional cases.

What you can check

A dash means not measured, not zero.

Public workbench
OpenFictional scenarios, fixed responses
Live-assessment report
—None yet
Probe set and scoring method
—Not published

What’s real, and what isn’t claimed

Real today

  • The public workbench: example scenarios, responses and findings you can open and inspect.

Not claimed

  • The demo does not score your system, run live AI or evaluate a prompt you type.
  • A passed code review is not live-assessment acceptance. No live-assessment report, customer or date yet.
  • The probe set and scoring method are not published.

Questions

What is the intended deliverable?

A scored report designed for an outside reader to check. The public workbench shows what example findings look like.

Is it available?

The demo is open to anyone. Live assessment is still in research; you can ask about its scope through Contact.

How does it relate to JCR?

They are related independent tools. PROBE looks at how an AI system behaves under adversarial input; JCR is a checkpoint, in active development, that allows or denies what an agent may do, or holds it for a person to approve, and records a receipt for each decision.

Other tools

In active development

JCR

Checkpoint and receipts for agent actions

Explore JCR Beta

RADAR

AI platform changes, explained, with corrections

Explore RADAR Public alpha

CCP Study Lab

Learn to investigate a security claim

Explore CCP Study Lab

Discuss the evidence you need.

The same discipline behind these tools, applied to your systems.

Start an advisory conversationAll tools →