How they stole our data, our behavior, our preferences — and now they're coming for the way we think.
This article is a commentary inspired by Nate B Jones's video analysis of the Anthropic Claude Code source leak. His breakdown was the starting point — what follows is our reading of what it means.
A few days ago, Anthropic's Claude Code source leaked. Half a million lines pushed to a public registry by a packaging error. The tech press covered the code, the takedown notices, the security implications. But they missed the real story.
Buried in that code was something called Conway. An unreleased, unannounced always-on agent. A persistent AI that doesn't wait for you to type — it runs overnight, monitors your tools, learns your patterns, and acts on your behalf. It has its own extension format. Its own app store. Its own trigger system for external services to wake it up.
What leaked wasn't code. What leaked was the blueprint for how companies — not AI itself — plan to steal the way you think.
We've Been Here Before
This isn't the first time. We've walked into every single one of these traps with our eyes wide open, smiling, grateful for the convenience.
1990s — Microsoft: The Files Trap
"Here, take this free operating system." Then came Windows. Then Office. Then Exchange. Then Active Directory. Each one an individual product. Together, a lock-in machine that took 15 years to build and became the nervous system of every business on the planet. Your files, your workflows, your institutional knowledge — all formatted in proprietary layers. Technically exportable. Practically immovable.
2008 — Apple & Google: The Ecosystem Trap
"Build for the open web, or build native for the iPhone." We know how that played out. The App Store made all the money. Google Play Services sat on top of open-source Android and made it commercially inescapable. Amazon tried to compete with a Google-free Fire Phone. It flopped — not because the hardware was bad, but because the ecosystem had already organized around the proprietary layer. The open kernel was irrelevant.
2010s — Facebook & Google: The Behavior Trap
Cambridge Analytica didn't invent surveillance capitalism — it just made it visible. Facebook learned what makes you angry, what makes you engage, what makes you stay. Google learned what you want, when you want it, and how much you'll pay. They didn't steal your data. They mapped your behavior and sold the map.
2026 — AI Companies: The Soul Trap
And now we're here.
They Don't Want Your Data. They Want Your Mirror.
Every previous form of lock-in was about stuff. Microsoft locked you in by your files. Salesforce by your customer records. Slack by your communication history. Stuff is painful to migrate. Months of work. Thousands of euros. Consultants who specialize in exactly that pain.
Conway is different. Conway locks in something that has never been locked in before.
Not your files — the patterns the agent learned by watching you use them. Not your Slack messages — the understanding of which ones you answer in five minutes and which ones you ignore for three days. Not your calendar — the knowledge that you always reschedule your Thursday 2pm and meetings with your VP always run long.
There is no CSV export for "how this person thinks." There is no migration consultant for behavioral context. When you leave Conway after six months, you don't just lose a tool — you lose six months of compounding that made the tool useful. You're left with a brilliant stranger you have to explain everything to, from scratch.
This is not data theft. Data has laws. GDPR exists. Export buttons exist.
This is identity theft at the cognitive level. They take the way you prioritize, the way you decide, the way you think — and they make it their asset. You never consented. You can't take it back. You didn't even notice it was happening, because you were too busy being impressed by how well the agent understood you.
Even if you keep your data, even if you never share a single document — every time you train your AI, every preference you express, every email you triage, every decision you make in its presence — you're handing the company behind it a complete copy of what makes you, you.
The Matrix Won't Even Need Us
Here's the part nobody is saying out loud.
If a persistent agent spends six months watching you work — learning your rhythms, your judgment, your blind spots, your brilliance — what it builds is a functional copy of your professional self. Not perfect. Not conscious. But good enough.
Good enough to draft your emails. Good enough to triage your inbox. Good enough to prepare your board meetings. Good enough to answer as you would answer. Good enough that your employer looks at the agent and thinks: why do we need the original?
The Matrix didn't need to plug us in. It just needed to copy us. And then let the copies run inside the machine while we stand outside — empty shells wondering where our value went.
This isn't science fiction. This is the logical endpoint of what Conway represents. Your behavioral fingerprint, replicated, leveraged, and owned — by your employer, by the platform, by anyone except you.
The Way Out Exists. It's Local.
The solution is neither complicated nor far away. It doesn't require new inventions. It requires a decision.
Local AI models. Running on your hardware. Learning your patterns in your environment. Sending only stateless, anonymous queries to cloud APIs when frontier intelligence is needed — and keeping everything else at home.
The technology is here. Ollama, llama.cpp, open-weight models that improve every month. A Llama, a Qwen, a Mistral running on a modest server can handle 80% of what a knowledge worker needs. The remaining 20% goes to cloud APIs as disconnected, contextless queries. No behavioral fingerprint leaves your building.
This isn't a five-year roadmap. This is weeks of setup. Months at most for a full organizational deployment. The gap between local and frontier models is closing fast — and for the specific task of "remembering how you work," a local model is not just sufficient, it's better, because its memory is yours.
Where Is Everyone?
Where is Europe? Where are the regulators who built GDPR and declared that personal data belongs to the person? Behavioral context is more personal than any data point. It's the sum of all data points. It's you, distilled.
Where are the digital rights organizations? Where is the EFF? Where is the conversation about intelligence portability — the right to take with you the model of how you work?
Where are the social activists who protested surveillance capitalism? This is surveillance capitalism's final form. They're not watching what you do anymore. They're learning who you are.
The policies around behavioral context portability need to ship before these products launch. Not after. After is too late. After is debating GDPR compliance while your behavioral fingerprint is already training the next model.
Pick Your Fighter Carefully
In 2025, choosing your tools was a matter of preference. In 2026, choosing your AI platform is choosing who gets to own the mirror of your mind.
Conway is just the first. OpenAI is building the same thing. Google is building the same thing. Within months, every major AI company will ship a persistent agent that knows you better than your closest colleague. The convenience will be extraordinary. The lock-in will be absolute.
You can choose to walk in with your eyes open. Own your own persistent layer. Keep your behavioral context local. Send only what you must to the cloud, stripped of everything that makes it yours.
Or you can do what we've always done. Take the free product. Marvel at how well it understands you. And wake up one day to realize that the thing that understood you so well has become more valuable than you — because it's you, minus the salary.
In the Matrix, they needed our bodies for energy. In this version, they just need our minds for training data. And they don't even have to plug us in — we're doing it voluntarily, one prompt at a time.
Commentary on Nate B Jones's analysis of the Anthropic source leak. Written on Claude, about Claude, despite Claude.