In the space of just a few days, the conversation around AI has shifted. The focus is no longer simply on how fast these systems can be deployed, but on how to control, govern, and constrain systems that are already approaching the limits of their design assumptions. AI is moving into the control layer, where systems act, coordinate, and increasingly influence outcomes. That changes the risk from isolated model error to systemic interaction risk. Here’s what I learnt this week on AI in Trading:

1. Anthropic’s Mythos Signals a Step-Change in Cyber Capability

Anthropic’s announcement of Claude Mythos Preview, deployed through its Project Glasswing initiative, highlights a new frontier in AI capability. Frontier models are now capable of identifying and exploiting software vulnerabilities at a level comparable to highly skilled experts, but at far greater scale. This materially lowers the cost and effort required for offensive cyber activity, but also potentially offers powerful defensive capabilities if deployed responsibly. Rather than making the model broadly available, Anthropic has opted to work with a closed group of major infrastructure and technology firms to apply the model in controlled environments – Amazon Web Services (AWS), Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, The Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Read more here: https://www.anthropic.com/glasswing

Why this matters for trading: The risk from AI is no longer limited to incorrect signals or model error. As AI systems move into execution workflows and infrastructure layers, the risk becomes systemic – as highlighted frequently by Jon Danielsson – https://www.linkedin.com/in/jon-danielsson/ – potentially affecting routing, allocation, and hedging decisions. When the control layer becomes partially autonomous, this shifts firm-level model risk issue to a market structure and coordination problem.

2. Regulators Are Stepping Up

This raises the risk of widespread exploitation across critical infrastructure – financial systems, energy grids, healthcare, and government networks. Unsurprisingly governments are rapidly assessing the risks involved with talks underway in both US and UK – https://www.reuters.com/world/uk/uk-financial-regulators-rush-assess-risks-anthropics-latest-ai-model-ft-reports-2026-04-12/. At the same time, there is growing recognition that these capabilities can also be used defensively – to identify vulnerabilities, improve software quality, and strengthen resilience across both proprietary and open-source systems.

Why this matters for trading: Regulation is beginning to shift from conceptual AI governance towards greater operational oversight – including testing, auditability, and lifecycle controls. This marks a transition toward treating AI not as an innovation layer, but as core financial infrastructure requiring greater supervision – as we have seen since the start of the year from the MAS guidelines – with new UK considerations to introduce standardised testing for bank AI models in play which mark an acceleration in regulatory thinking – read more here: https://www.ft.com/content/3053b547-5e55-4520-9b95-828c417a5d79

3. Retail Signals the Speed of Adoption

AI-driven automation is also advancing rapidly in retail investing. Brokerage platforms are beginning to introduce AI-assisted workflows, and there is increasing experimentation with systems that support portfolio construction, rebalancing, and risk management. Last week Public rolled out of AI brokerage agents (read more here – https://www.mindfulmarkets.ai/ai-in-trading-2026-from-co-pilots-to-control-layers/), this week moves retail automation further from concept to deployment with parallel experiments now running fully autonomous portfolios with real capital, handling everything from research to rebalancing – read more here https://www.linkedin.com/pulse/claude-now-running-50k-portfolio-zero-human-override-anyone-beli%C5%ABnas-cloaf/. While fully autonomous portfolio management remains at an early stage, the direction is clear: decision-making is gradually being delegated to systems rather than tools.

What this means for trading: Markets today are not designed for correlated, always-on, machine-driven participants. As agentic systems scale, order flow may become more synchronised, more reactive, and more concentrated in time. This increases the potential for feedback loops, herd behaviour, and intraday instability, particularly during periods of stress.

4. Cisco Targets the Agent Security Layer

At RSA Conference 2026, Cisco introduced DefenseClaw, a secure agent framework designed to embed security directly into the development and runtime lifecycle of AI systems. Alongside broader initiatives in identity management, Zero Trust access, and AI-enabled security operations, Cisco is focusing on ensuring that agents are secure by design, not just monitored after deployment. Read more: https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m03/cisco-reimagines-security-for-the-agentic-workforce.html

What this means for trading: Execution risk is becoming architecture-dependent, governance is becoming programmable, and control over agents is emerging as a competitive differentiator. At the same time, increased trading volumes are exposing fragilities in the underlying infrastructure – particularly around cloud concentration, hybrid complexity, and operational dependencies. As AI scales, these weaknesses are likely to be amplified rather than reduced.

5. AI Infrastructure as a Strategic Consideration

The infrastructure supporting AI – data centres, compute capacity, and energy supply – is becoming more visible and strategically important. The rapid expansion of hyperscale infrastructure, combined with geographic concentration and energy dependency, introduces new dimensions of risk – not just applicable to governments and industry – but also now the City.

What this means for trading: Market infrastructure is now indirectly coupled to compute availability, geographic distribution, and energy resilience. This creates new forms of concentration risk and potential single points of failure. At the same time, there are early signs of diversification, with regions such as Europe developing broader AI ecosystems across models, infrastructure, and applications. This diversification may help reduce systemic concentration over time.

Markets are transitioning from human-controlled workflows to machine-driven systems faster than the control frameworks are evolving. The challenge is no longer building AI – it will be governing it at scale. This is where industry coordination becomes critical. The FIX Trading Community is actively working on extending existing standards into the AI domain – from enhanced, machine-readable execution tagging to the evolution of algo certification and now looking at lifecycle governance, including testing, auditability, and transparency.

As AI agents begin to drive execution, shape liquidity, and interact across venues, the absence of common standards risks fragmentation, opacity, and systemic instability. The next meeting scheduled for May 7th, contact me direct or the FIX Program office for more information.

Thanks again for reading – more to follow next week. As always let me know your feedback/comments

Many thanks

Rebecca