The report
A misleading market headline need not move a price very far by itself to become dangerous when many automated traders respond in the same direction. FARSIGHT tests that proposition in a simulated market: a fabricated SEC-style headline had about a 0.1% direct effect without the agent swarm, compared with roughly 9.6% mean peak drawdown in the study's multi-agent condition. The numbers depend on the modeled order book and participating agents. Headline experiment.
Another set of experiments used an automatic liquidation rule. The authors report 400 of 450 simulated shocks crossing a 50% collapse threshold. The rule creates correlated selling pressure after a trigger, which can deepen the fall that caused the trigger. It is a useful illustration of reflexive automation, but not an estimate of how frequently live markets will crash. Liquidation tests.
The two experiments should not be compressed into a headline saying an AI agent caused a real flash crash. The study uses a synthetic environment and academic trading-agent frameworks; broker-level execution rules, circuit breakers, market depth and nonagent participants could change the result. Research design.
The mechanism has two checkpoints. First, a low-trust news item becomes trading evidence without independent provenance or contradiction checks. Second, independent-looking agents share similar policies and liquidation thresholds, making their responses correlated. Defenses therefore have to address both source validation and the portfolio-wide consequences of actions that would each appear acceptable in isolation. No published result here validates a particular intervention against all real-market conditions.