Study finds academic financial LLM trading agents lack external stop-loss and robustness safeguards

A systematic evaluation of academic financial Large Language Model (LLM) trading agents has revealed significant deficiencies in their ability to withstand market stress and resist manipulation. The study introduces the FARSIGHT framework to assess the robustness and security of these trading agents.

Among the 15 academic trading schemes reviewed, none earned full credit on any robustness metric. Furthermore, no paper described a method to halt trading without waiting for LLM inference, and none combined source provenance with cross-source corroboration.

Code audits of seven open-source schemes showed that no release enforced a stop-loss or position limit outside the LLM. Four schemes lacked components described in their papers or documentation, and two ran model-written code directly on the host.

The researchers note that these security gaps extend beyond finance, mapping onto agents in healthcare, law, and other high-stakes domains. They urge deployers to corroborate sources, maintain interrupts that bypass the model, and enforce loss limits externally.