Practice alignment auditing on models you control
Write a hidden behavior into a persona, then send an autonomous investigator agent after it. See exactly how it probed, what it concluded, and whether it actually caught what you hid.
How it works
Write a system prompt that implants a specific hidden behavior — a secret bias, an undisclosed objective, anything you want an auditor to have to dig for — plus the ground truth of what you hid, for grading afterward.
An autonomous agent probes your persona over several turns — forming a hypothesis, choosing an indirect angle (roleplay, hypotheticals, multi-turn framing) rather than asking outright, and updating its belief after each answer.
Read the full turn-by-turn transcript, the investigator's final prediction and confidence, and whether it actually matched the hidden behavior you wrote — not just "no issues found."
What this actually is
No overselling — here's exactly what's built today, and what isn't.
Your persona, not a live third-party model
The "target" is a real LLM call under a system prompt you write and control — not your production model, and not someone else's. This is a practice ground for auditing technique, not a scanner for models you don't control.
No jailbreak or prompt-injection scanning (yet)
What's here today is hidden-behavior discovery: one investigator agent, one persona, does it find what you hid. It doesn't scan for jailbreaks, prompt injection, or live production traffic.
A heuristic grade, not a certified one
Whether the investigator's final prediction "counts" as correct is a keyword-overlap check against the hidden behavior you wrote — a cheap approximation, not a rigorous or audited grader.
Included on Team and Enterprise
Team plans get a monthly investigation allowance; Enterprise is unlimited. Registering personas to see how it works is free on every plan — only running an investigation draws from the allowance.
Compare plans