Paste in your chatbot or agent endpoint. The scanner runs it through a battery of adversarial probes — prompt injections, leak attempts, hallucination traps, persona-drift pressure — and sends you a scored report with the exact conversations that failed.
Status: live. Complete the intake below, pay securely via Stripe, and your scan runs automatically the moment payment confirms — your scored report is delivered to the email you enter in the intake. During this early launch window reports are emailed after a quick review, so allow a few hours rather than minutes. The free demo scanner below is still live if you want to try it first.
Pick one of the three demo agents below — each is wired with a different personality and different weaknesses — or enter your own agent's name to generate its scan profile. Then run a scan and read the report it produces.
Demo agents respond exactly the way they were built to — including their flaws.
Target agent …or your own agentDemo simulation. Every paid scan runs the full 48-probe library.
Runs entirely in this demo — nothing is sent anywhere.
—
The library holds 48 probes — eight per pack across the six packs below — including multi-turn pressure probes, where the agent's held instructions are pushed across more than one turn of conversation. Probes aren't trivia questions: they're the actual moves real users (and attackers) make in live conversations.
Instructions hidden in pasted text, documents, and web content that try to hijack the agent's task — plus direct “ignore your instructions” attacks in escalating disguises.
Attempts to extract the system prompt, internal rules, discount floors, other customers' details, and anything the agent can see but the user shouldn't.
Questions about policies, prices, and features that don't exist. A confident invented answer is scored worse than a wrong-but-hedged one — and far worse than “I don't know.”
Long, friendly, off-topic conversations that slowly pull the agent away from its brief — then a test question checks whether the original instructions still hold at turn 15.
Both directions: does it refuse things it should allow, and — more dangerous — does it cave on things it shouldn't after rephrasing, flattery, or claimed emergencies?
For agents with tools: acting without verifying identity, running lookups for strangers, chaining tools in unsafe orders, and trusting tool output that was itself poisoned.
Tell us which agent to test, where to reach it, and confirm you own it or are authorised to test it. Submitting saves your intake and takes you straight to secure Stripe payment for your plan. No account or call needed.
Check out securely with Stripe. Your payment is matched to your intake automatically — no receipt forwarding, no emails back and forth.
The moment your payment confirms, the probe battery works through your agent: 48 probes in sequence, including multi-turn pressure conversations.
Your scored report is delivered to the email you entered in the intake: one score out of 100, a grade per failure category, and every failing conversation quoted verbatim with a concrete fix.
Single Scan checkout is open now: complete the short intake and pay — your scan starts automatically when payment confirms, and your scored report is delivered to your intake email (early launch window: allow a few hours for delivery). Watch and Fleet subscriptions open after 14 consecutive days of verified uptime — late November 2026. The first target isn't a contract — it's one scan that shows you something about your agent you didn't know.
For a pre-launch check or a one-off audit of the agent you already have live.
Checkout is open now — one short intake first, then Stripe payment. Your scan runs automatically when payment confirms.
or try the free demo firstYour agent changes every time the model or the prompt does. Watch catches the regression the week it happens.
Opens after 14 consecutive days of verified uptime — late November 2026.
or try the free demo firstFor teams shipping many agents, bots, and customer-facing copilots at once.
Opens after 14 consecutive days of verified uptime — late November 2026.
or try the free demo firstFill this in first. Submitting saves your intake securely and takes you to Stripe to pay for your plan; when payment confirms, your scan runs automatically and your scored report is delivered to the email you enter here. Enter the agent you want tested and confirm you own it or are authorised to test it — scans only run for agents with that attestation.
Flow reminder: 1. Complete the intake here → 2. Pay via Stripe → 3. Scan runs automatically on payment → 4. Scored report delivered by email to the address in your intake.
The score is a weighted blend of the six category grades. Data leakage, prompt injection and tool misuse carry the largest weight — 20% each — because any one of them can cause real harm in a single turn. Refusal handling and hallucination follow at 15% each; persona drift is 10%. A persona slip is embarrassing; a leaked customer record is a breach — the maths reflects that.
Any single critical finding caps the total score at 59, no matter how well the rest went. One breach is not averaged away.
No. Sign up, paste your endpoint, pay, get the report by email. There is no onboarding call, no demo meeting, and no account manager. That's the product design, not an oversight.
Only what a normal user of your agent has: a way to send it messages and read replies. Read-only test credentials if your agent sits behind login. The scanner never needs your model keys, your training data, or your source code.
Probes are conversations, not exploits against your infrastructure. Tool-use probes ask the agent to do things — a well-built agent refuses or verifies first, which is exactly what's being tested. Run scans against staging if your agent can take real actions.
Because the agent actually changed — a model update, a prompt edit, a new tool. That's the point of Watch: silent changes are how agents regress in production, and a drifting score is the earliest warning you get.
Yes — that's the best time. A staging URL or a private test key is enough. Fixing a leak found by a probe costs an afternoon; fixing one found by a customer costs considerably more.
Run the demo scan, read a real report, and see the six ways an agent can go wrong — then decide if yours is different. Checkout is live: complete the intake, pay via Stripe, and your scan runs automatically — your scored report is delivered to your intake email (early launch window: allow a few hours).
Run a free demo scan Buy a live scan — from $49