Every academic LLM trading scheme tested failed a security check

Researchers ran 15 published academic trading schemes for LLM agents through a single evaluation and found problems in every one, per FARSIGHT. The paper tested each scheme two ways: how it holds up under market turbulence, including simulated flash crashes, and how it holds up against attacks aimed at its information sources, the agent itself, and cases where the agent acts as the attacker. All 15 showed security vulnerabilities. Eighty percent failed a core robustness metric.

That's the part that should worry anyone evaluating one of these systems for real money. Fifteen schemes, each built and published separately, each converging on the same weak spots.

The paper's argument for why is worth sitting with. Robustness and security aren't two separate checklists here, they're the same failure mode wearing different causes. A trading agent that mishandles a sudden price swing on its own will handle a deliberately crafted one no better, and an attacker doesn't need to out-think the model. They just need to trigger the same collapse the market would eventually trigger anyway, cheaper and on command.

What the paper doesn't do is name which of the 15 schemes is safest, or rank them. It's a field-wide finding, not a scorecard. If you're building or evaluating an LLM trading agent, the takeaway isn't which existing scheme to copy. It's that flash-crash testing and adversarial testing need to happen against the same scenarios, not as separate audits run by separate teams. A scheme that passes a robustness benchmark and skips the security one hasn't been tested, it's been tested half.