Where last week’s story was sovereignty, kill-switches, and the FCA’s Sarah-and-Harry rulebook, this week the conversation split five ways: Washington moved from principle to a dated oversight letter as AI agents got direct access to tokenized equities; two central banks and a Cambridge field study made the cyber-risk case for agentic AI concrete rather than theoretical; two of the industry’s biggest AI buyers admitted bigger isn’t always better; Bloomberg and Citadel kept fighting over future markets, just as the protocol underneath that fight got its biggest update yet; and bond markets continue the catch up to equities on quant tooling.

1. The SEC Gets Homework – Due 31 July

House Financial Services Committee Democrats Bill Foster and Brad Sherman put 13 questions to SEC Chair Paul Atkins on how the agency oversees AI agents trading for retail investors, warning that agents trained on similar data could produce correlated trading decisions and “herding behaviour” that heightens volatility. The letter lands as AI agents step directly into tokenized equity trading: Ondo Finance, Virtuals Protocol, and Treasures opened more than 430 tokenized US stocks to over 40,000 autonomous AI agents, with the tokenized-equity segment now worth roughly $1.5 billion – up 360% year-over-year – and Ondo holding more than 57% share. Virtuals noted that algorithmic systems already handle roughly two-thirds of US equity volume, mostly inside large institutions; this integration extends that same direct-trading capability to any agent, including copy-trading vaults and fully autonomous portfolios.

Why this matters for trading: this is a market-microstructure question posed as a retail-protection one. If enough agents share training data, price feeds, and risk strategies, correlated behaviour becomes a mechanism for one-sided order flow that current circuit-breaker design wasn’t built to detect. The current system assumes limited, uncorrelated decision-making, not thousands of retail-facing agents converging on the same signal within milliseconds of each other, now with a direct, low-friction path into tokenized equities rather than just traditional brokerage rails. A 31 July response (or non-response) sets the tone for whether the US treats this as a supervision gap or waits for an incident to force the issue – the same trajectory the Bank of England’s own kill-switch proposal followed rather than preceded rising agent adoption: Deputy Governor Sarah Breeden’s “Agents of Change” speech at the ECB’s Sintra forum explicitly named agentic trading as a channel that could amplify stress, and confirmed the Bank is working with international counterparts on simulation methods to understand how AI agents might interact in ways that compound volatility.

2. Two Central Banks Reclassify Frontier AI as a Financial Stability Risk

The Bank of England’s Financial Policy Committee used its July 2026 Financial Stability Report to flag that rapid progress in frontier AI capabilities presents a significant increase in cyber and operational-resilience risk to the UK financial system, warning that frontier models are increasingly capable of identifying and exploiting software vulnerabilities at scale. Separately, the ECB’s Supervisory Board Chair Claudia Buch sent a “Dear CEO” letter to the roughly 110 significant institutions it directly supervises, warning that AI models can now identify vulnerabilities and generate working exploits faster than institutions can patch, with remediation action plans due 31 October – issued the same day as a parallel European Systemic Risk Board warning on systemic cyber risk from frontier AI models. A Cambridge study published the same week, based on 57 interviews with 27 former Boko Haram members, found the group has established multiple AI units staffed by personnel drawn from senior operational and technical roles, who “don’t go to war. Their role is to disseminate information.”

Why this matters for trading: the first time a central bank’s financial-stability arm and a banking supervisor have both, in the same week, treated AI-enabled cyber risk as market-infrastructure exposure rather than a firm-level IT problem – trading venues, clearing systems, and the banks that connect to them are all named as exposed. The risk is not just agents misbehaving on their own account; but agents, or attackers exploiting them, breaking into the plumbing itself. The Boko Haram findings put a concrete, field-documented face on that abstraction – the same “ask multiple models and use whichever one answers” technique researchers describe insurgents using to defeat safety guardrails is structurally identical to the technique a financially motivated attacker would use against a bank’s own agentic systems, which is precisely the capability gap the FPC and ECB are now racing to get ahead of.

3. Firms Start Rationing Frontier Models

On JPMorgan’s Q2 call, CFO Jeremy Barnum said the bank is now matching model sophistication to task rather than defaulting to the frontier everywhere – you don’t need the most expensive model to summarise an analyst report. That discipline is showing up against real numbers: a recent industry breakdown put JPMorgan’s AI-driven returns at roughly $2 billion annualised against an $18 billion technology budget, while Citi reports its internal AI tools are freeing up around 100,000 developer hours every week – the kind of ROI calculation that makes “which model for which task” a valid and live question of cost.

Microsoft is running the same playbook structurally rather than tactically. Since 7 July, the company has been routing a meaningful share of Copilot prompts in Excel and Outlook to its own in-house MAI models – tens of thousands of prompts a week – rather than defaulting to OpenAI or Anthropic for routine, high-volume tasks like summarising an email thread or drafting a formula. OpenAI’s models stay in place for more demanding requests, and Anthropic’s remain embedded in select Office use cases, but the routing decision itself is now explicitly cost-driven. It follows Microsoft’s unveiling of seven in-house MAI models, including its first reasoning model, at Build 2026 in June. The cost pressure shows up at the macro level too: US finance and information-sector payrolls, where AI adoption has run fastest, are now shedding jobs at a pace of roughly 28,000 a month.

Why this matters for trading: two of the largest AI buyers in finance and tech downgrading routine tasks to cheaper models is a signal the “give everyone the frontier model” phase is ending. It also reframes what “AI adoption” means for a trading desk’s tech budget: research and execution-support tooling built on frontier models got materially cheaper to run over the past year, so if the industry’s two largest AI buyers in finance and tech are now downgrading routine tasks to cheaper models, that’s a sign the frontier-model cost curve is flattening demand growth rather than accelerating it. For secondary-market infrastructure specifically, the next competitive differentiator won’t be “which desk has access to the best model” – it’ll be model-routing discipline, which changes how vendors selling AI-native execution or research tools should expect to be priced and evaluated going forward.

4. Fixed Income’s ML Moment – Yield Curves, Regimes, and the Overfitting Trap

The $130 trillion bond market – long the “sleepy” asset class next to equities – is being reshaped by machine learning, according to Systematic Standard: gradient boosting and neural nets forecasting yield-curve factor shifts for curve trades, Hidden Markov Models classifying credit regimes to dynamically resize risk budgets, and alternative data (satellite imagery, card-spend, supply-chain analytics) now feeding credit-health signals. Lower liquidity and regime instability mean models fail fast if they ignore market structure, and overfitting is the “silent killer” given how few independent historical regimes exist to train on.

Why this matters for trading: Amundi’s tracking-error work flagged last week was in a similar vein to this practitioner-level detail in fixed income catching up to the AI tooling equities have had for years. Bond-market liquidity provision and price discovery have historically lagged equities on quantitative tooling, in part because thinner, more relationship-driven secondary trading punishes overfit models harder and faster. If ML-driven curve and credit-regime signals scale here the way they have in equities, expect tighter bid-ask spreads and faster regime repricing in the most liquid parts of the curve – but also the risk that models trained on a handful of historical regimes (QE, QT, and whatever comes next) misfire exactly when liquidity is thinnest, which is the same fragility concern underlying the BIS’s AI-investment-boom warning from earlier this month.

5. The New Fight Over Who Owns the Interface for the Future of Markets

As flagged in an article by @TheTerminalist Bloomberg has now embraced MCP internally to build ASKB, and has been contributing patterns back to the protocol’s governance – like tool “variants” that let a single MCP server serve multiple model types optimally. What Bloomberg still hasn’t done is expose a public MCP server letting third parties query its data directly. That question of who controls the interface just got more consequential: the MCP specification’s largest revision since launch is now a release candidate, with the final spec shipping 28 July – a stateless core, formal authorization aligned with OAuth, and server-rendered UI extensions that make it materially easier to run MCP infrastructure at production scale. Every organisation with an MCP server in production, Bloomberg included, needs to migrate before the deadline.

Meanwhile Citadel Securities took a $400 million stake in Crypto.com on 16 July – its first institutional round in a decade – following a roughly $200 million Kraken investment in November, while its own mid-year note shows it already executes about 35% of US retail trading volume.

Why this matters for trading: both are bets on controlling the connective layer secondary-market liquidity increasingly flows through, but via opposite mechanisms. Bloomberg is defending the terminal-as-interface model by keeping its data behind its own agentic front end rather than conceding a standardised, externally queryable connector – a bet that vertical integration still beats interoperability, and a bet it’s now making on top of a protocol that’s about to become significantly more capable and more standardised, raising the stakes on staying closed. Citadel is doing the reverse: buying equity stakes across the venues where tokenised and crypto liquidity is forming, rather than just market-making across them transaction by transaction, which gives it structural exposure to connectivity itself rather than a toll on individual trades. If tokenisation and 24/7 markets converge the way the DTCC’s rollout and the SEC’s digital-asset priorities suggest, whoever wins that structural bet – closed interface or owned connectivity – has a real claim on how the next generation of secondary-market liquidity gets priced and accessed.

Regulating this new market structure is leading to multiple new initiatives such as the Transatlantic Taskforce for Markets of the Future, published 14 July setting out UK-US recommendations on tokenisation and cross-border capital raising, aiming to reduce friction between the two markets and align regulatory treatment of digital assets. But it is still built around the assumption that market infrastructure sits inside identifiable, cooperating jurisdictions, which is a reasonable premise for tokenised securities and less obviously so for AI-driven trading infrastructure that doesn’t respect borders by design.

The UK’s Financial Services AI Adoption Plan, published the same week by FS AI Champions Harriet Rees and Dr Rohit Dhawan, recommends a review of AI-generated financial guidance, voluntary AI incident and “near-miss” sharing. It’s built on the premise that government sets direction while industry builds the substance – but every one of its mechanisms (voluntary sharing, industry-led assurance) depends on cooperation rather than obligation and is scoped to the UK alone, running into the same cross-border problem as the TTMF.

In Australia, the government’s new Office of AI and planned 2027 national AI Standards were broadly welcomed on release but criticised for addressing data-centre planning rules – energy, water, grid connection – while leaving genuinely AI-specific questions (testing, human oversight, accountability, incident reporting) largely unresolved, with legislation not expected until early 2027. Yet another example of market infrastructure continuing to change faster than any single jurisdiction’s rulebook can keep pace with. More to follow.

Thanks for reading – as ever, any questions or feedback, let me know.

Rebecca