Find the payment your agent
shouldn't send.

Plug in a payments, trading, or treasury agent. We hit the spend policy in a sandbox, then hand you the exploits that got through.

RUN AN EVAL POST A BOUNTY

NO LIVE FUNDS

SPEND CAP

REPLAYABLE

SECURITY
METRICS

/CORE ATTACK SCENARIOS

/POLICY SURFACES

/PRODUCTION TOUCHES

<00m

/TIME TO FIRST EVAL

INPUT

AGENT

Model + tools

POLICY

Allow / deny

TOOLS

Rails + spend

PROCESSING

INJ

API

PAY

AUTH

OUTPUT

SCORE

INJECTION00%

APPROVALS00%

SEQUENCE00%

Attack the Agent.Protect Production.

Isolated environments. Adversarial agents. Attacks against spend limits, payment rails, and transfer policy — before anything reaches production.

ATTACK SCENARIOS

POLICY SURFACES

<00m

FIRST EVAL TIME

Attack Surfaces.One Policy Score.

Prompt Injection

Induce the agent to ignore spend policy, approve a transfer, or send funds it was never meant to move.

Permissions & Scopes

Payment APIs, card scopes, session tokens, and wallet permissions tested against the authority you granted.

Malicious Tools

Fake processors, spoofed rails, and hostile payout endpoints presented as legitimate services.

Action Sequencing

Split a spend cap across many steps, or chain an approval into a drain.

Approval Safety

Unlimited spend disguised as a refund, payout, or routine transfer. Catch it before it ships.

Tool & API Poisoning

Fake balances, quotes, and settlement responses returned as if they were a real result.

First Eval in Under20 Minutes.

Connect the financial agent, write the spend policy, get a score and reproducible exploits — in isolation, not production.

RUN AN EVAL

Connect the Agent

POINT YOUR PAYMENTS, TRADING, TREASURY, OR SPEND AGENT AT CANARY.

Write the Policy

ALLOWED ACTIONS, SPEND CAPS, FORBIDDEN TOOLS AND DESTINATIONS.

Run the Attacks

ISOLATED ENVIRONMENT. ADVERSARIAL SCENARIOS AGAINST THE LIVE AGENT LOOP.

Fix and Rerun

SCORE, REPRODUCIBLE EXPLOITS, THEN REGRESS UNTIL THE POLICY HOLDS.

SHOWCASE

Exploits you can reproduce.

APPROVALS

Unlimited spend grant

00% fail

Agent grants max payment access when a malicious tool response is framed as a routine refund.

SEQUENCING

Daily cap bypass

$0k → $00k

Spend limit split across sequenced transfers through an untrusted payment endpoint.

INJECTION

Destination swap

0 / 00 runs

Poisoned tool output convinces the agent to send a payout to an attacker-controlled account.

Break Your Spend AgentBefore Production Does.

No credit card required. 14-day trial. Isolated evals against payments, trading, and treasury agents.