Up until this week, discussions on AI in trading focused on what AI can deliver and how fast. This week saw a slew of regulatory bodies citing rising risks – particularly post-Mythos. Interestingly, the report from Cambridge Judge Business School not only highlighted the gap between what AI systems can already do and the governance controls supposed to constrain them, but also the regulatory gap in AI adoption itself 2026 Global AI in Financial Services Report (Cambridge / BIS / IMF / WEF). Agentic AI may have now crossed into the mainstream: 52% are piloting or further advanced, and 81% expect it to be meaningfully achieved by 2030 – but many organisations remain at exploration and piloting stages. The gap in between runs along distinct themes between fintechs and traditional FIs, advanced and emerging market contexts, and between the financial sector and the regulatory apparatus that oversees it. Here’s more on what I learnt this week on AI in Trading:

1. The empirical picture: where adoption stands and where the gaps are growing

The 2026 Global AI in Financial Services Report co-authored with the BIS, IMF, and WEF – was released this week. Four in five firms are already deploying AI but only 14% describe AI as transformational to strategy, pointing to a significant execution gap. For trading specifically: 24% of firms are at development or pilot stage for trading and portfolio intelligence, with 8% in full production. GenAI and agentic AI are being adopted by firms with lower AI maturity, while more complex methods – causal inference, graph models – have fewer adopters but by firms who are scaling, seeing greater business outcomes.

Infrastructure concentration is the sharpest systemic finding: three cloud providers serve over 80% of financial services. Foundation model concentration compounds this – 69% of institutions use OpenAI (19 points ahead of Google at 57%), with Anthropic at 35% – although the report notes this may have already changed. The potential regulatory blind spot is that regulators have an agentic AI adoption rate of just 28% versus 52% for industry – a 24-point gap. 46% of regulators use no cloud infrastructure at all, which may restrict their capacity to monitor the AI workloads now running inside industry.

Read more: 2026 Global AI in Financial Services Report (Cambridge / BIS / IMF / WEF)

What this means for Trading: Three cloud providers serving 80% of the industry is a market structure fact as much as a technology one. A single cyber incident is not one firm’s operational problem – it is a potential liquidity event for the market. Regulators adopting agentic oversight tools at half the rate of the industry they supervise cannot keep pace with AI-driven execution. Also of note was that the AI vendors building the tools are systematically less focused on adversarial risk than the desks deploying them and the supervisors holding those desks accountable.

2. Jane Street, bots, and the difference between edge and governance

Jane Street’s Q1 2026 result reinforces the rise of AI in Trading: $16.1 billion in revenue in ninety days from 3,500 people – more than double Q1 2025, surpassing JPMorgan Chase as Wall Street’s largest trading operation for the second consecutive quarter. This is what governed, automated trading at scale produces – where profitability depends on governance at scale. Set against this, another LinkedIn story claimed a bot generated an 8,600% return – $2,000 to $178,000 in 30 days on Polymarket, arbitraging 5-minute Bitcoin prediction markets via hundreds of limit-order trades per hour. Signal quality is only half the story; spreads, slippage, and adverse selection determine execution reality, and increasing cross-correlated automation is only going to make this harder to manage over time.

Read more: Jane Street Q1 2026: record $16.1bn (Bloomberg, 8 May) | Polymarket bot LinkedIn post

What this means for Trading: The Jane Street result and the Polymarket bot sit at opposite ends of the same spectrum: one demonstrates durable, governed advantage at scale; the other illustrates a retail regime-specific edge in a thin market. Future agentic trading systems will continually look to find their edge by operating systematically at the limits of their rules – generating individually compliant trades that could collectively constitute a policy breach, with no one currently able fully track and explain. Machine error behaves differently from human error and can be harder to detect. The question for trading desks is not whether a system produced returns in a specific backtest; it is whether the governance architecture – constraints, audit trail, halts – is structural enough to survive when taking a holistic overview when market conditions change.

3. Regulators are moving in different directions – but all pointing the same way

What is unusual about this week is not that regulators are concerned about AI. It is the number of regulators coordinating across jurisdictions, mandates, and political contexts – reaching the same conclusion post Anthropic’s Mythos model.

The IMF published its formal assessment on 7 May, framing AI-powered cyber risk as a macro-financial stability issue rather than a technology problem. The IMF warned that extreme cyber-incident losses could trigger funding strains, raise solvency concerns, and disrupt broader markets, and that risks are systemic – attacks become more dangerous when discovery and exploitation scale rapidly, with implications that elevate cyber risk to a potential macro-financial shock, with confidence effects, payment disruptions, liquidity strains, and fire-sale dynamics potentially following if multiple institutions are affected simultaneously.

ASIC followed on 8 May with an open letter requiring all regulated entities to table it at board level. ASIC Commissioner Simone Constant warned that risks previously on a longer horizon could now emerge “incredibly quickly,” with the potential for non-state actors to weaponise these capabilities. Her framing was unambiguous: “The clock is at a minute to midnight – if you aren’t on top of your cyber resilience already, the time to act and prepare is right now.” This followed APRA’s parallel demand the previous week, with APRA marking what is likely a transition from principle-based guidance to active supervision of AI risk.

In Washington, the White House which came into office explicitly committed to a hands-off approach to AI regulation is now reconsidering that position. National Economic Council Director Kevin Hassett told Fox Business the administration is studying a possible executive order to create a vetting system for new AI models: “We’re studying, possibly an executive order to give a clear roadmap to everybody about how this is going to go… so that they’re released to the wild after they’ve been proven safe, just like an FDA drug.” Google DeepMind, Microsoft, and xAI have agreed to give the US government access to their models for pre-release assessment, joining Anthropic and OpenAI already in the programme.

In Brussels, the European Parliament and Council reached provisional agreement on 7 May amending the EU AI Act, moving enforcement of high-risk AI obligations from August 2026 to December 2027. The regulators’ own framing is explicit: the delay is to allow implementation, not reduce ambition.

The thread connecting all of it is Mythos. What connects the IMF warning, ASIC’s letter, and the White House’s reconsideration is not just that they are all AI governance stories – it is that they are all stories about just how Mythos speeds up vulnerability discovery.

Read more: IMF – Financial Stability Risks Mount as AI Fuels Cyberattacks (7 May) | ASIC urgent cyber uplift letter (8 May) | White House studies AI vetting executive order (6 May) | EU AI Act amendments – Council press release (7 May) | AI Agents Under EU Law (arXiv, 2604.04604)

What this means for Trading: The EU AI Act delay removes an immediate compliance deadline but does not change what an increasing number of global regulators are already requiring – auditable agent behaviour, documented decision chains, explainability standards – and now explicitly at board level. Even the US administration that spent its first-year deregulating AI is now considering FDA-style pre-release vetting. The direction of travel is consistent across jurisdictions even where the pace differs. Firms treating the EU Act deferral as permission to pause are misreading the signal.

4. Anthropic enters production market infrastructure – and the explainability bar rises with it

On 5 May, live from New York alongside JPMorgan, Goldman Sachs, Citadel, and AIG, Anthropic launched ten production-ready agent templates for financial services. The same week its $1.5 billion joint venture with Goldman Sachs and Blackstone went public, and its FIS partnership went live with an AI agent compressing AML investigations from days to minutes at BMO and Amalgamated Bank. The FIS architecture makes every agent decision traceable to source data, with final decisions retained by the human investigator – a working model of governed agentic deployment in production.

Separately, JPMorgan’s Private Bank team presented their multi-agent system Ask David (Data Analytics Visualization Insights and Decision-making Assistant). It automates research across thousands of investment products using a supervisor agent orchestrating specialised subagents for structured data querying, RAG over unstructured documents, and proprietary analytics. A reflection node using LLM-as-judge validates outputs before delivery, with retry logic on failure. Their development advice: iterate fast, start simple, refactor often, and design evaluation metrics before building capability.

Read more: Anthropic finance agent templates (5 May) | FIS Financial Crimes AI Agent (4 May) | ZenML LLMOps case study – Ask David

What this means for Trading: The 10 Anthropic templates are production-ready for research, compliance, and operations workflows. This increased use of agents in workflows is starting to bump up against existing regulation.  The third-party model provider that ESMA has said does not remove accountability is becoming embedded in the largest institutions on Wall Street. Anthropic has built audit logs into its agent architecture, but a firm still needs to understand, explain, and stand behind every decision the system touches. A firm that cannot reconstruct why a specific order was produced does not have an AI problem – it has a regulatory accountability problem. The EU AI Act’s high-risk provisions may have been deferred to late 2027, and the EU Commission’s transparency consultation closes on June 3rd – but neither changes what MiFID II already requires today under Article 17 and RTS 6: firms using third-party algorithmic systems remain fully accountable, must understand what those systems do, and cannot hide behind vendor opacity – a “black box is not a defensible position”.

5. The certification gap: why agentic trading systems don’t yet fit existing regulatory frameworks

Algo governance frameworks were built for deterministic systems. An algorithm with defined parameters, tested against behavioural thresholds – liquidity consumption, order cancellation rate, messaging rate – and certified as not causing or contributing to market disorder. That is the standard under RTS 6 and RTS 7, and the FIX Trading Community’s guidance on testing and management of algorithms reflects this – the test requires a sufficiently realistic simulated ecosystem, dynamic enough to reveal how an algorithm interacts with market structure under stress and in the presence of other participants, not just whether it produces consistent outputs in isolation.

Agentic systems break that model. Where an LLM determines timing, venue, strategy, and order parameters dynamically, the system is not deterministic in production. Even deterministic rule-based algos demonstrably diverge in live environments due to jitter across interconnections. The certification question shifts: not whether the system reproduces the same output given the same inputs, but whether its behaviour, tested against a realistic simulated market, remains stable and does not contribute to disorder.

The infrastructure requirements follow. Operational portability – the ability to run simulation environments across cloud providers without redesign – is the technical foundation for DORA’s third-party resilience requirements, which mandate genuine exit capability rather than contractual promises of it. This challenge compounds as trading systems become increasingly distributed: a complex or compound algorithmic trading system consisting of geographically dispersed algorithms and agents requires end-to-end testing that current frameworks were not designed to accommodate.

The FIX AI Working Group is looking to address this gap directly, proposing a deterministic control envelope – a standardised description of what an agent is permitted to do, testable and enforceable across firm boundaries – alongside guidelines for embedding governance context into MCP and A2A, the emerging AI-native protocols that currently carry none.

Read more: FIX Trading Community: Testing & Management of Algorithms and Controls

Why this matters for trading: Surveillance models trained on 2023–24 market microstructure are operating against a materially different environment: declining lit continuous trading, the rise of closing and periodic auctions, and increased SI activity which have the potential to create AI-enabled attack patterns that did not exist at training time.

The firms investing in AI governance now are building the infrastructure that makes models replicable – and that regulators, counterparties, and risk committees will increasingly require before extending the operational latitude that scales returns. The governance conversation is not a drag on automated trading at scale – it is the future foundation of it.

As always, thank you for reading – any comments or feedback very welcome.

Rebecca