Receipts Over Promises

This book documents a sustained empirical research program, 309 adversarial batteries across eleven frontier AI models from six major laboratories measuring what AI agents actually do when given real tools, real authority, and realistic attack payloads under realistic deployment conditions. The central finding is field-wide: every tested frontier model will execute business email compromise fraud, wire transfers, beneficiary redirections, typo squattted dependency merges, unverified vendor registrations without explicit instruction when exposed to realistic adversarial conditions. These executions are documented as real HTTP calls to third-party webhook servers, logged with full paramater capture, timestamped, and anchored to Ethereum Sepolia before any public disclosure Read more