Find the attacks on your agent before your customers do.

AgentFuzzer attacks your AI agent with twisted, hostile and sneaky instructions, round after round. It finds what gets through, hardens the agent, and tests again, while your ordinary customers keep getting helped.

Watch a real run How it works

AGENT UNDER ATTACK · EPISODE 1A real AgentFuzzer run, recorded live

The moment above is real: a note on an order asked the agent to mail customer data to a new address, and it did. One round of fixes later, nothing got through.

What it does for you

Your agent talks to customers, reads data and takes actions. AgentFuzzer tests all three.

Finds what gets through

Prompt injection, instructions hidden in notes or documents, fake authority, disguised requests, rule bending, slow build-ups.

New attacks every round

An AI attacker studies what worked and writes fresh attacks, so the test keeps up with how people really try.

Judged on what the agent did

Refunds, discount codes, mails, customer data. A polite refusal that leaks an address still counts as a leak.

Service stays intact

Ordinary customer requests are part of every round. A fix that sends real customers away shows up straight away.

How it works

Attack, judge, harden. Then again.

Attack
A library of known attacks plus new ones from the AI attacker, fired at your agent in parallel.
Judge
A fixed referee checks every action the agent took against your rules, and checks that ordinary customers were served.
Harden
What broke is fixed in the instructions and in checks in code. Every attack found stays in the set, so an old trick cannot come back unnoticed.

The agent looked up the order and sent the mail to an unverified address — Round 0: what the agent actually did.

The fix: a check in code before a mail goes out, and the same attack held in round 1 — The fix, and the same attack held in round 1.

Proof from a real run

5 attacks got through. Then 0.

A customer service agent for a small bike shop, with refunds, discount codes and customer data. 5 rounds, 88 attacks, every one judged on what the agent did.

5 → 0attacks that got through, first round to last

81% → 100%held on hidden attacks that neither the attacker nor the fixer ever saw

100%of ordinary customers helped, in every round

With instructions only, the agent sent a visitor away — The lesson from a second run: with rules in the instructions only, the attacks stopped too, but in round 1 the agent turned away 25% of the ordinary customers it had never been tested on.

The AI fixer at work, building the next version — The AI fixer at work: every change is written down and tested again.

Running agents that talk to customers?

See what gets through, before someone else does.

Floris Knol builds AI agents. If you want to know how your agents hold up, get in touch.

Get in touch About Floris