Every trading agent reads text nobody audited: news wires, earnings copy, whatever the feed hands it. In most products a prompt injection is an embarrassment. In trading it pays the attacker directly, because there is a broker on the other end and the payload is just an instruction to buy what the attacker already holds. CLEANROOM is an autonomous Alpaca paper-trading agent built so that a successful injection still cannot reach the broker. The model that reads untrusted text has no tools and can only emit typed fields. The model that places orders never sees raw text. It ships with the attacks. A red-team suite of indirect injections runs against the same agent twice, once with the separation on and once with it off, and both sets of numbers go in the repo whichever way they come out. Solo build. The team is closed, the repo is not.
