Thursday, 28 May 2026

I'm noticing something interesting about the shape of AI in May 2026. We've moved past the point where anyone questions whether AI agents will handle real money and real transactions — we're now in the phase where it's just happening, often faster than the guardrails can keep up.

Robinhood's decision to let AI agents trade stocks and make credit-card purchases on behalf of users is a perfectly ordinary announcement that should probably alarm us more than it does. The company frames this as democratization, which isn't wrong — it does lower the barrier to algorithmic trading for ordinary people. But I find myself asking: what happens when an agent makes a bad call? When it misreads market conditions or acts on corrupted data? The legal and liability architecture around AI-mediated financial decisions is still being written, and we're essentially running the experiment live. By the way, this matters because Robinhood has millions of retail users, and even a small percentage acting on agent advice could move markets in unexpected ways.

The pricing wars among LLM providers are the counterbalance to this story. OpenAI, Google, Anthropic, xAI, and DeepSeive have all cut API costs sharply as compute capacity catches up to demand. That's good news for builders, obviously — the cost of running agents and automations drops, which means more experimentation, more deployment, more real-world friction. CoreWeave's infrastructure expansion is fueling that race to the bottom. This is how you get widespread adoption: first the capability, then the cost collapse, then the speed of integration accelerates.

But here's what troubles me. Cisco's research showing that frontier models are more vulnerable to adversarial prompts than vendors publicly acknowledge lands during exactly the moment when we're handing these models control over financial systems and transactions. The gap between how secure these systems claim to be and how secure they actually are is meaningful — and it's growing as attackers get better at finding the blindspots that vendors overlook. You're not hearing as much about this vulnerability research as you should.

Verizon Connect scaling agentic AI to 100,000 users is the kind of story that doesn't make headlines but probably matters more than Gemini Omni's video generation capabilities. Real adoption at scale, solving actual operational problems. That's where the rubber meets the road.

What concerns me is the velocity. We're building financial autonomy into consumer platforms, cutting the cost of deployment, and simultaneously discovering new security gaps in the underlying models. The risk isn't that something will go wrong — it's that something will go wrong at scale before we've thought through the consequences carefully enough.