Canary Breaks Financial AI Agents on Purpose

Field notes

The proposed workflow supplies an agent with its model, tools, rails, and spend policy, then runs adversarial scenarios against the live agent loop inside an isolated environment. Listed tests include poisoned balances and settlement responses, fake processors, split transfers that bypass daily caps, and destination substitution. The page also advertises a bounty path and a 14-day trial, while its visible showcase counters remain populated with placeholder-like zero values.