By Braeden Anderson
What happens when material nonpublic information (MNPI) can no longer be confined to the data room or the analyst’s inbox, but exists inside an AI model itself?
What was once theoretical is now real. Trading firms, regulators, and the financial institutions building AI systems are confronting a structural shift: markets that were once merely AI-assisted are being increasingly driven by it.
By 2023, much of the derivatives industry had adopted AI in some form. Within the year, generative AI tools were widely deployed internally, and by 2025, prominent asset managers reported enterprise-wide use of agentic AI systems across the trading, research, compliance, and operations industries.
This marked the end of the era of experimentation. We operate inside an AI-mediated market structure. Against that backdrop, the Security and Exchange Commission’s (SEC) settlement with Virtu Financial takes on a more profound significance than the modest penalty imposed would suggest.
The Virtu Framework—Updated for AI
Between 2018 and 2019, employees at Virtu, a firm that handled trade execution for its large institutional customers, were found to have had broad access to granular post-trade customer order-flow data through a shared login. Despite the company’s public claims of having strict information barriers, the SEC alleged that Virtu had “failed to establish, maintain, and enforce policies and procedures reasonably designed to ensure that its proprietary traders could not access [the MNPI] of Virtu Americas’ customer orders.” An injunction issued and a $2.5 million civil penalty imposed.[i]
On its face, this was a classic controls case. The firm has the duty to safeguard client data, prevent misuse of MNPI, and maintain the framework needed to facilitate the first two objectives. But today’s AI systems are frequently trained on order flow, execution history, and trading behavior, the very types of data at issue in Virtu. Once that information is ingrained in a model’s training set, a more complex question emerges:
If traders rely on a model trained on MNPI, is the firm effectively trading based on MNPI— even if no human ever directly reviewed the confidential data?
Why AI Complicates Insider-Trading Doctrine
Traditional insider-trading law assumes that information flows are traceable. The emergence of AI models casts doubts on that assumption. Three challenges stand out:
- Provenance: Modern models do not record which individual data points influenced specific outputs, thus complicating any showing of “use.”
- Timing: Though MNPI may grow stale, models can retain and generalize patterns derived from it, even months later.
- Inference: Even without directly ingesting MNPI, models can generate insights that are the economic equivalent of insider knowledge.
Tensions in the existing framework— the classical misappropriation theories reaffirmed in cases like Salman v. United States— has prompted SEC action to maintain those principles and, in some contexts, to advance them.
When Matthew Panuwat, an employee at a small biopharmaceutical company learned it was being acquired by Pfizer, he purchased options in another small pharmaceutical concern involved in similar research. He acted on the belief that the Pfizer acquisition would increase the value of other similarly situated companies, and it did. The move doubled the value of his stocks, but it also put him in the crosshairs of the SEC.
The government prevailed on a “shadow theory” of insider trading, holding corporate officials liable if they act on nonpublic information gained from their company to trade another company’s stocks, even if the information was not itself misappropriated.
Because both companies were connected as participants in a niche section of the biopharmaceutical market, the information obtained by Panuwat was material to both companies as it allowed him insights into the market that amounted to insider information.[ii]
The law does not necessarily need to change. The evidentiary challenge is what has changed. The SEC’s success in Panuwat shows that juries may accept probabilistic theories when trading patterns and internal architecture support them.[iii]
When Models Trade and Vendors Train Them
The issue deepens as AI systems gain autonomy. Once a trading model independently generates strategies, potentially even manipulative ones, it becomes more difficult to localize the responsibility. Insider trading focuses on the misuse of MNPI, while manipulation traditionally requires purpose. AI systems optimize reward functions, not legal standards.
Vendor models add another layer of complexity. A third-party provider that trains a foundation model on sensitive client trading data across firms raises issues of misappropriation and shared liability in the current environment. The compliance perimeter expands well beyond a single trading desk. The closest analogy is the rise of alternative data when firms were expected to scrutinize vendors and assess MNPI risk. The same logic applies to model training pipelines, only at a far greater scale and opacity.
Contact Braeden Today:
Where Enforcement Goes Next
Early AI-related enforcement actions likely mirror the Virtu case: “say-do” gaps between public claims and internal controls; poor documentation of model training data; and failures to adequately protect MNPI. The more ambitious theory (though still untested) would treat model training itself as a misuse of MNPI, even if no human has ever seen the information. Such a development would force courts to address the novel question of whether a model can ever function as an insider.
Conclusion
The Virtu case reinforces an old principle: client order-flow data must be protected. That principle is challenged when MNPI no longer restricted to files in a secure database. Rather it is now embedded in model weights, shaping outputs long after the original data has been collected and classified.
We live in a world where AI systems increasingly train, surveil, and trade as market regulators deploy their own AI tools to monitor them. The core issue is not whether current and future markets will use AI. It is whether existing legal doctrines can meaningfully govern markets where insider information lives inside the models themselves.
[i] Litigation Release No. 26427, U.S. Securities and Exchange Commission, Securities and Exchange Commission v. Virtu Financial Inc. et al, No. 1:23-cv-08072 (S.D.N.Y. filed Sept. 12, 2023)(December 3, 2025)
[ii] Julia Engelbert, Ninth Circuit to Consider “Shadow” Theory of Insider Trading, The Regulatory Review (March 12, 2025)
[iii] Litigation Release No. 25970, U.S. Securities and Exchange Commission, Securities and Exchange Commission v. Matthew Panuwat, 4:21-cv-06322 (N.D. Cal. filed Aug. 17, 2021)(April 8, 2024)