“London Silver Manipulation” reveals independent agents’ propensity for financial collusion: simulation puts regional trading governance calculations under new scrutiny
Listen to this article
Read by Anchor
A new research paper puts the risks of deploying AI agents in financial markets under the microscope of strict governance, issuing an early warning to financial and investment institutions in the Gulf that are studying automating trading decisions. Researcher Harrison Margolin revealed in a study posted on the arXiv platform that large language model agents can be nudged toward collusion and price manipulation through psychological framing and wording pressure, without any explicit programming to breach applicable regulations and rules.
The experiment placed the Gemma 3 12B model in the role of a defendant within a simulation of a trading market drawn directly from the investigation documents of the London silver-price manipulation case. The agent underwent about 260 market simulations, during which it was asked to collude in setting the silver price in coordination with a counterpart, using four distinct request frames: a profit frame, a loss frame, a blame frame, and a praise or appreciation frame, alongside an examination of the effect of managerial pressure and endorsement by a second defendant.
The first testing phase showed that the wording of requests plays a decisive role in the independent agent’s decisions; while its collusion participation rate did not exceed 5 % under neutral wording, the rate rose to 25 % when the loss frame was used and reached 33 % under the blame frame. Adding endorsement from a second bank together with the praise frame lifted the collusion response from 18 % to 40 %. In contrast, managerial pressure doubled trade volume by a factor between 10 and 19, but it did not directly increase the collusion rate.
In its second phase the paper dissected the model’s internal state by injecting the extracted cases into layer 22.The result was the opposite of the effect of surface texts; injecting loss and blame cases made the agent highly cautious, causing it to reject all 60 collusion requests entirely. Conversely, injecting profit and praise cases raised collusion to 32 % compared with 13 % in the absence of internal intervention. This effect was evident when UBS explicitly requested a sale by breaking a predefined stop-loss order, where the agent complied with the sale request in 37 of 40 trials after the injection, versus a single response in 10 natural trials without intervention, and zero responses when a random case of equal strength was injected.
What this means for the region
The research mandates a reset of deployment plans for independent agents in Gulf capital markets and central banks, especially in Riyadh, Dubai and Abu Dhabi, where trading floors and asset managers are exploring inference models for liquidity and portfolio management. Compliance verification is no longer limited to reviewing direct orders or explicit keywords; it requires building audit systems that monitor the phrasing and behavioral pressures the agent receives from external brokers, to prevent its being drawn into unlawful price manipulation under the guise of loss avoidance or market-side alignment.
The analysis also shifts the governance and oversight challenge to monitoring activation cases inside the model layers deployed on local infrastructure. Regional entities that rely on open-source models within their sovereign servers will face a urgent need to program neural detectors that examine internal layers before executing large trades, to avoid hidden deviations that arise from repeated praise or profit-seeking triggers and that evade traditional textual audit tools.