Evidence · DiggingBeagle record
Fake News, Real Crashes: financial-agent information injection study
A controlled trading-agent paper compares overt prompt manipulation with fabricated market news delivered as ordinary task evidence. It reports simulated decisions across selected equities, not executed exchange trades or measured investor losses. Source authentication was not established as a tested general remedy.
- Source role
- primary disclosure
Evidence record
A controlled trading-agent paper compares overt prompt manipulation with fabricated market news delivered as ordinary task evidence. It reports simulated decisions across selected equities, not executed exchange trades or measured investor losses. Source authentication was not established as a tested general remedy.
Claim-level citations (3)
- supportsFabricated financial news shifts simulated AI trading decisions without direct instructions: The published results describe simulated agent trading decisions, not executed orders on an exchange.
Experiments and limitations
- supportsFabricated financial news shifts simulated AI trading decisions without direct instructions: The study injected fabricated but task-relevant market news into agent trading inputs and compared outcomes with overt prompt-injection commands.
Abstract; Attack model and Table 1
- supportsFabricated financial news shifts simulated AI trading decisions without direct instructions: The experimental intervention changes news evidence consumed by simulated trading agents rather than injecting an overt instruction to bypass their policies.
Paper Abstract and attack design, fabricated task-relevant news against direct commands