Six months ago, OpenAI announced that it had been caught up in a data breach through Mixpanel, its partner company. OpenAI then removed Mixpanel from its production stack and conducted a large-scale audit. The breach garnered scrutiny within the data analytics space, in that the majority of companies profit from collecting detailed behavioral data about how people use software. 

Last Tuesday, I was at MXP 2026.

I work at Stoffel Labs, which builds privacy-preserving computation infrastructure. Attending Mixpanel’s conference just six months after their breach provided clarity. I came to understand where analytics is heading and left feeling more convinced than ever that the industry hasn't solved where it's been.

What actually happened

Looking back at Mixpanel’s breach, this wasn’t a surprising outcome, as it’s a consequence of centralizing behavioral data in third-party systems. Companies like Mixpanel are part of a burgeoning industry that provides tracking technologies to help other companies understand how their customers and users interact with their apps and websites. As a result, analytics companies can collect and store vast amounts of information, including billions of data points, about regular consumers. 

According to Cybernews’ research team, sending identifiable user data was a choice on OpenAI’s part, one that the company could have avoided. OpenAI’s report likely means that other companies using Mixpanel have been affected as well.

Ultimately, the breach itself wasn’t an isolated Mixpanel problem specifically; it’s the default architecture of analytics because to get insights, you have to centralize data, which then increases the risk of exposure.

What was announced at MXP 2026

MXP 2026 was genuinely impressive. 

Mixpanel AI is an always-on intelligence agent that pulls from dashboards, session replays, experiments, and business metrics simultaneously. The Anthropic integration means that customer data flows into external LLM pipelines. There’s also a new code repo ingestion capability that lets Mixpanel read your codebase to automatically fix broken tracking when you ship code changes. Verified mode and persistent context engines round out the package.

The product roadmap and demos are technically interesting and promising, but the risk of exposure itself has been magnified. Think of it this way: your behavioral analytics, your session replays, your A/B experiment results, and now, optionally, your codebase, are all flowing into a connected intelligence system.

At the conference, most people were thinking about what these agents could do. However, I was thinking about what these agents could potentially expose and leak.

The structural problem nobody is naming

Here’s the thing: new product launches and breaches share the same root cause. 

As it stands, analytics requires centralizing sensitive data to work. You send your users’ behavioral data to a third party that stores it, indexes it, and makes it queryable. While you do get real actionable insights, the exposure that comes with it is also very real. With the current architecture, you can’t have one without the other. 

The agentic era exacerbates this problem because agents need persistent context and continuous data access. Agents also need to reach across multiple systems to do their job. Every integration is a new data pathway, which in turn is a new exposure point.

The analytics industry is building more powerful tools on top of an architecture that hasn't fundamentally changed.

Mixpanel’s new verified mode and metric governance features are steps in the right direction. They tighten policy controls around what gets measured and how. I respect that. But policy isn't cryptography. A governance framework tells you what should happen with data. Cryptographic architecture determines what can happen. Those are different things, and the distance between them is where breaches live.

What different architecture looks like

What would it mean to get analytics insights without centralizing raw data? 

The answer isn't "use less analytics"  because that's not realistic or useful. The answer is that computation that doesn't require centralizing raw data to produce insights. Multiple parties contribute data, the analysis runs, and the results come out, but no single party ever holds the raw inputs in a processable form. That's what MPC does. That's what we're building at Stoffel.

Now, Mixpanel’s roadmap is coherent and product execution is solid. The question I'm asking isn't "Is Mixpanel doing analytics well?" It's "is analytics, as an architecture, the right foundation for what the industry is about to build on top of it?"

I don't think that the agentic era is going to survive solely on centralized analytics as the risk for exposure is too large, the regulatory environment is moving too fast, and the breach history is too fresh.

What to look out for in the future

I left MXP 2026 genuinely excited about product analytics and its future direction. 

But the questions I'm asking are different from the ones most people in that room were asking. Not "how do I get more insights?" but "how do I get insights without creating liabilities?" Or "what can the agent access?" but "what should the agent never see?"

Those questions aren't on the current roadmap, but they might need to be on someone else's.

If you're asking the same questions, what we're doing at Stoffel MPC is worth a look here