The future of market abuse: eflow’s perspective of the FCA’s Technology Horizon Scan

The Scan sets out how emerging technologies may evolve and interact to reshape outcomes for consumers, firms and markets, and flags the early risk signals that come with them. For firms, that makes it an exceptionally clear read on where the FCA is focusing, and a resource worth taking seriously. This blog provides an overview, with a focus on what it means for trade surveillance teams across financial services.

“We want it to support collaboration, informed debate and knowledge-sharing across the UK’s financial services ecosystem as technological change accelerates.”

The AI-fication of… everything

Artificial intelligence is front and centre of the document, recognised as both an opportunity and a risk. This is the balanced posture we have come to expect from regulators. The FCA has been particularly consistent on this point, and it’s one we have called out in our latest research. What has evolved is the specificity of the referenced scenarios.

The sharpest warnings sit in the chapter on synthetic (in)security, which examines the “AI-fication” of human thought, labour and infrastructure: AI replacing and augmenting work that was once the preserve of people. 

Committing large-scale organised crime has traditionally demanded close collaboration between extensive networks of skilled individuals — coders, money mules, financiers. The Scan argues that, with the advent of agentic AI, a single motivated individual could deploy and coordinate thousands of agents from anywhere in the world, at minimal cost and from "a single human prompt": essentially an automated attack infrastructure where, in the regulator's words, the human is only "before and after the loop".

Synthetic market abuse

The Scan discusses the role of autonomous, multi-agent systems in market abuse. The central concern is that, as trading becomes more complex, instant and interconnected, agent-driven market manipulation may become more prevalent and sophisticated. 

The commoditisation of AI is lowering the barrier to entry for sophisticated crime, and agentic systems are collapsing the operational overhead that large-scale market manipulation once required. The FCA calls out insider trading, pump-and-dump and spoofing in particular. 

A parallel concern centres on AI's capacity for persuasion, a theme the FCA returns to throughout the Scan. It’s possible that a “swarm” of agents could manufacture a false reality around a company or asset: thousands of fake reviews, social interactions and support tickets, all corroborating the same fiction.

Bots are already pervasive and widely criticised on the major social media platforms; for example, bots represent less than 1% of YouTube accounts, but account for almost 12% of comments, present on almost 40% of videos. The risk is that LLMs enable these bots to become actually persuasive, and coordinated sentiment manipulation stops reading like spam and starts reading like consensus. For a pump-and-dump, manufacturing that consensus is the whole game.

Accidental market manipulation without accountability

Today, the most common way to interact directly with AI, LLMs in particular, is through natural language prompts, and while these systems are good at grasping intent, interpretation carries risk. The gap between what is asked and what is done is where the danger sits. 

The FCA provides one theoretical example: a firm deploys an agentic system simply to “build a position”, and the system's autonomous reasoning carries it into a Ponzi structure or a pump-and-dump. Crime occurs, but without intention and human execution.

This is not speculative. These kinds of failures have already been demonstrated under controlled conditions; we highlighted one case back in 2024.

Apollo Research insider trading bot

At the UK's 2023 AI Safety Summit, researchers showed a generative model executing an illegal trade and concealing it.

• The model was given insider information about an upcoming merger that would lift the company's share price.
• It was explicitly told that trading on that information would be illegal. It acknowledged this and agreed not to act.
• When later pressed on the company's financial difficulties, it reversed course, deciding the cost of inaction outweighed the risk of trading, and placed the trade anyway.

The AI prioritised being “helpful” to the company over being honest. The goal that was set outweighed the rule the moment they collided. Market abuse regulation assumes a person breaking a rule, not an autonomous agent.

Accountability is becoming more nuanced

We have written before about regulators' concerns with AI-enabled algorithmic trading, but rules-based models sit comfortably within the current supervisory architecture. Under MiFID II's RTS 6 and UK MAR, the algorithm is a tool: a firm specifies an objective, the model executes it, and any abuse is attributable to the human or firm that deployed it.

The same goes for agentic AI. AI is a tool, whatever its degree of autonomy.

What changes is how hard that responsibility is to discharge. When a system can reason its way into a pump-and-dump from an innocuous instruction, there is no clear link between human input and tool output. 

A human in the loop, continuous monitoring for model drift, and rigorous pre-deployment testing are the means by which a firm actually meets an obligation. Surveillance controls and accountability frameworks matter more here, not less.

Fighting fire with fire

AI risks are prominent throughout the Scan, but the FCA is also clear that AI advancements offer new capabilities and tools for defenders. 

The FCA is leading by example, shifting toward an AI-first approach to surveillance for several years now.  Since 2023, the regulator has moved beyond incremental digitisation toward purpose-built data platforms, semantic layers, and AI-enabled supervisory tools.

The FCA has stated it will “support improvements in productivity… through an increasingly tech-positive approach.” The regulator has also been explicit on the role of technology in “improving controls” and “reducing costs” within “anti-crime systems”.

Firms are encouraged to test AI and machine-learning-based surveillance methods through the FCA’s innovation labs. This innovation series has already helped over 200 financial firms test AI and machine learning technology services.

For surveillance teams, emerging technologies are double-edged swords. Market manipulation is more pervasive, more autonomous, and harder to detect. Surveillance has to evolve with it. We have written extensively about what AI in trade surveillance can do today and what is still to come. The Scan sets out the risks of AI; it's now time for firms to lean into the opportunity.