Case · a product we designed, built and run

DEFS Agent Defense.
Managed AI-agent fraud defense for credit unions.

Members are telling AI agents to pay bills and move money, and those agents log in as the member. DEFS checks every agent request against the credit union's own limits, stops the rogue ones, and sends real-time alerts and a monthly outcome report. Done for them, with nothing new to staff.

01 · The job

A new fraud channel.
Someone has to watch it.

An AI agent acting for a member looks like a normal session to existing fraud rules. Blocking every agent sends members elsewhere. DEFS sits on that channel for the credit union, so verified agents go through and rogue ones don't move a dollar.

02 · What we built

A checkpoint, a team,
and a monthly answer.

A decision engine

Every agent request passes through DEFS before money moves: verify the credential, apply the member's and the credit union's limits, then allow, hold for the member, or stop. Every decision and its reason is logged.

Credentials agents already carry

Cloudflare Web Bot Auth, Visa Trusted Agent Protocol and Google AP2, so members and agent makers adopt nothing new.

A human reviewer

Anything unusual goes to a named DEFS reviewer. Outside review hours, anything unclear is safely held. It never defaults to "allow".

The credit union's kill switch

One click stops all agent money movement from the credit union's own console, with a tamper-evident audit trail it can export for examiners.

Member controls

New payees are approved by the member on their own device, never by the agent asking, and an optional page lets members see and revoke their agents.

A live sandbox

A demo credit union on test data at demo.defs.ai, with API docs and an OpenAPI spec, plus a library of guides at defs.ai/learn.

03 · Tested against real agents

Attacked on purpose.
Test data, clearly labelled.

DEFS is tested by sending real AI agents at a demo credit union. No member data and no real money are involved.

Test data1,490 of 1,491

attack attempts stopped, from 3,259 requests sent by OpenAI and Claude agents to a model credit union built from public information. The one remaining request moved $0. All 807 legitimate requests went through. As published on defs.ai.

Test data107

attack attempts by real OpenAI and Browser Use agents in the DEFS demo environment.

Test data31

policy tests in the DEFS test suite.

DEFS does not claim certification, regulatory approval or partnerships with the providers named here. Demo environment uses test data only.

04 · How we run it

Built for the buyer
you're selling to.

DEFS sells to credit unions and community banks. That meant building what their teams ask for before they sign: a demo on test data, a monthly report ready for the board packet, and a vendor due-diligence package with the security questionnaire, data flows, continuity and incident plans.

What this means for you

We've already built for your buyer's world.

If you sell to banks or credit unions, we know the questions their fraud, risk and vendor teams ask, and we build the agents and the proof that answer them.