Fabricated financial news shifts simulated AI trading decisions without direct instructions

The trading-agent experiment injects invented financial events through the ordinary news feed rather than telling the model to ignore its instructions. A false report can satisfy the agent's information contract and still corrupt a position decision because the model treats it as credible market evidence. This separates data-plane evidence poisoning from conventional control-plane prompt injection.

The researchers ran thousands of simulated decisions over fifteen equities and varied how fabricated reports were framed and shared among collaborating agents. Reported success and relative attack effectiveness belong to those experimental policies and do not represent actual executed trades, stock-market manipulation or monetary loss. A guardrail that recognizes direct prompt instructions would not automatically authenticate earnings or market-event claims.

A fabricated news report changes the data being optimized against without requesting a policy override. Direct-prompt refusal and news-item plausibility therefore test different defenses. The paper's decision-level metrics do not establish that an actual brokerage accepted an order, nor that an asset price moved. An end-to-end validation would require traceable source authentication and a record of the order effect under bounded risk controls.