Half of studied options-trading agents permit potentially unbounded downside

A model can suggest or submit an options order, while the broker interface decides whether it becomes financially binding. In the surveyed prototypes, some order-capable systems lacked a fixed independent loss ceiling or mandatory human approval. Short options exposures can exceed the amount originally committed when orders are not separately restricted.

The authors examine 26 options-trading systems after screening 765 repositories and other candidates, reporting 13 architectures with potentially unbounded downside and only 10 of 22 order-sending systems with independent hard limits. The sample is not representative of all deployed trading products, and the paper does not document 13 realized customer losses. Broker API features, repository maturity and simulated operations have different denominators.

Apply loss and order-size limits at the broker or signing layer using the exact instrument and positions, with separate evidence for authorization of each consequential action. Agent capability scores do not substitute for bounded financial exposure. The defect is the missing independent trade constraint. Primary evidence.

Additional evidence and trust boundary

A finance agent can remain correct about trade direction while an attacker influences expiration, strike, volatility or the number of option legs. The resulting exposure cannot be bounded by a prompt asking the model to be careful. Order execution must recalculate maximum loss with a separately trusted market feed and an independent risk engine before sending the broker request. The 13 potentially unbounded systems are configuration findings, not 13 realized investor losses. Supporting record.