Why this exists
Companies are giving AI agents authority over money faster than they are testing them for it. The platforms that build these agents test them with cooperative simulated customers. Security red teams test for jailbreaks and injected instructions. Almost nobody sends a counterparty who simply wants the money, and almost nobody reports the result in dollars. Meanwhile, tribunals and regulators treat what an agent says as what the business said.
What we publish now
The product is in development. In the meantime we publish the work it is built on: the attack library, the scenario library, a sourced incident tracker, regulation explainers, practical guides and a glossary. It is free to read and cite.
How we work
- Facts are sourced, examples are labeled, and corrections are noted on the page. See the methodology.
- We are independent of agent platforms and model providers.