A viral post claimed a $2.2M engineer leaked Anthropic’s living neural brain and got fired the same day. Almost every word of that is false. But the smartest debunk of it was also wrong.
Here is a sentence that got tens of millions of views this year:
“Anthropic’s lead engineer making $2.2M/year leaked the company’s internal Obsidian brain, and got fired the same day.”
It goes on to describe what was allegedly leaked: a live neural network with 21 inputs, ten-plus hidden layers, ReLU activation, thousands of connections firing in real time as the system processes every decision. First layer 64 neurons, second 37, then 22, all the way to outputs. “This is not a diagram,” the post insists. “This is the living brain of the company thinking right now.”
I want to be gentle here, because this went viral among smart people. But that paragraph is roughly equivalent to looking at a subway map and announcing that you have discovered a living circulatory system currently pumping blood through the city. It’s a picture of notes. It is not thinking. Nothing is firing. ReLU is not involved. The only thing activating in real time was the engagement algorithm.
So the obvious move is to write the debunk. And a very good one went around: it’s just a knowledge graph, the story isn’t magic, it’s discipline, this is what years of linking documents looks like. Sharp, sensible, true about knowledge graphs.
Except that debunk quietly accepted the premise. It assumed there was a leaked Obsidian brain to debunk.
There wasn’t.
And once you start pulling that thread, you find something far more interesting than a leak: a real, boring, technical accident with four separate fabrications stacked on top of it like sediment, each one added by a different person for a different reason, each one traveling further than the truth underneath it.
Let me take the layers off one at a time. It’s a genuinely instructive autopsy.
Layer zero: the thing that actually happened
Underneath everything, there is a real event, and it’s almost aggressively unglamorous.
In late March 2026, someone at Anthropic published a release of Claude Code to the public npm registry and accidentally included the source map files. Source maps are the little .map breadcrumbs that let a debugger trace minified production code back to the original source. Ship them publicly and you’ve effectively shipped your source. Roughly half a million lines of internal TypeScript, out in the open. A security researcher spotted it within hours. A post linking to it did over twenty million views.
Anthropic’s statement was, in essence, “this was a release packaging issue caused by human error, not a security breach,” and no customer data or credentials were exposed.
That’s the whole event. A build step, missing a guardrail, at a company shipping fast enough that Claude Code’s run-rate revenue had blown past $2.5 billion. Speed creates pressure, pressure creates shortcuts, shortcuts create gaps, and eventually a human walks into one. Every engineer reading this has shipped something they shouldn’t have. Mine was a .env file. I still think about it.
Note what’s absent from that story. No Obsidian. No brain. No neural network. No $2.2 million. And, crucially, nobody was fired.
Layer one: “an engineer confessed, and was fired”
Within a day, a post appeared: “I was fired from Anthropic today. I was the engineer responsible for shipping the latest claude-code npm package… I take full responsibility.” It was well-written, contrite, technically specific. It named Undercover Mode and Bypass Permissions Mode. It apologized “to the team and to Claude,” which is a beautiful touch.
It was fake. The author had never worked at Anthropic. The confession was a marketing stunt for his own product, and he dropped a discount code for it in the replies. Community Notes eventually flagged it. A second person posted a similar “I was fired, this was my first engineering job” confession as a joke, and that one traveled too.
So before we even reach the Obsidian nonsense, we have two strangers competing for the role of a fired engineer who does not exist.
And here’s the actual truth about the firing, which is lovely and almost nobody amplified: nobody was fired. Boris Cherny, who created Claude Code, said publicly that the person involved retained the company’s full trust, that this was a process failure anyone could have fallen into, and that the fix belonged in the pipeline, not in a person. That is a genuinely admirable response, the kind of thing Toyota built a manufacturing philosophy on and most companies still can’t manage. It’s the single most useful lesson in this whole saga.
It got a fraction of the attention of the fake confession. Naturally.
Layer two: the Obsidian brain
Now somebody looked at the leaked repo, saw internal documentation, and reached for a bigger story. Enter the “Obsidian Brain”: 8,893 nodes, 4,729 links, 9,000+ documents. Marginalia Collection. Glossary Backbone. Master Index. Numbers that sound like evidence because they have decimal places.
And then the framing: the company building the most powerful AI in the world uses an Obsidian graph to manage its own innovations, and now it’s all public. He made $2.2M/year to know this. You got it for free.
Take a breath. Even if a graph like that existed, it would be… fine? Normal? A large organization keeps notes, and one plausible way to keep notes is a linked-notes app. That’s not a leak of the company’s soul. That’s a company having a wiki.
Download the Medium app
But the “$2.2M engineer, fired same day, living brain” package isn’t describing anything real. It’s assembling a myth out of a genuine incident, an invented person, and a screenshot.
Layer three: calling a knowledge graph a neural network
This is the layer that makes me laugh and then makes me sad, because it reveals how few people who post confidently about AI can distinguish between two things that merely look alike.
A knowledge graph and a neural network both render as dots joined by lines. That is the entire extent of their similarity.
A knowledge graph is storage. The dots are documents. The lines are links you typed. It is a map of what you wrote and how you decided to connect it. It sits there. It does nothing. Open it tomorrow and it will still be sitting there, doing nothing, faithfully.
A neural network is computation. The dots are units that transform numbers. The lines carry weights learned from data. Feed it an input and activations propagate, layer by layer, producing an output that wasn’t in any of the dots. It fires. That’s the whole point.
Describing an Obsidian vault as “a live neural network with 21 inputs and ReLU activation, thousands of connections firing in real time as the system processes every decision” isn’t an exaggeration. It’s a category error dressed as expertise. It’s like reading a phone book aloud and calling it a conversation.
And this is the part I’d underline for anyone whose job now involves an AI narrative: visual complexity is not evidence of intelligence. A tangle looks smart. A galaxy of nodes looks like thought. Our brains are terrible at this and always have been, which is precisely why the post worked. The graph looked like a brain, so people read it as a brain, and 8,893 nodes felt like proof of something.
So who was right?
Here’s the twist I actually want to leave you with.
The debunk post, the sensible one, said: it’s a knowledge graph, not a neural network; the story isn’t magic, it’s discipline; this is what years of meticulously linking documents looks like; stop chasing shiny tools and build boring systems consistently.
Every word of that is good advice. It is also a careful correction of a photograph of a unicorn.
The debunker got the mechanism right and the fact wrong. They knew enough to spot the category error and not enough to ask whether the thing being miscategorized existed at all. Which is, honestly, the most relatable failure in this whole story, and I say that as someone who has confidently corrected a colleague’s architecture diagram before discovering the system had never been built.
That’s the shape of misinformation now. It doesn’t arrive as one big lie you can reject. It arrives as a true event, wrapped in a false detail, wrapped in an invented person, wrapped in a technical claim that’s wrong in a way only specialists catch. Each layer is plausible given the layer beneath it. You can be completely right about your layer and still be standing on sand.
And notice the incentive gradient running through the whole thing. The true story, a build script needed a guardrail, and the company chose not to fire anyone, is useful, humane, and profoundly unshareable. Every fabrication that got bolted onto it made it more dramatic and more spreadable. That’s not a coincidence. That’s the physics of the medium. Boring truths don’t have velocity, and in an attention market, velocity is fitness.
The discipline point survives, and it’s the right one
I don’t want to throw out the debunker’s insight just because their example was made-up, because it’s the most valuable thing in this entire mess.
The instinct to look at a dense graph and think “if I had a system like that, I’d be smart too” is the same instinct that makes people buy a new note-taking app every eleven months. The tool is not the thing. Obsidian is a text editor with backlinks. It has no opinions and grants no powers. Whatever value lives in a mature knowledge graph was put there by someone who, on nine thousand separate boring occasions, decided to write the note and make the link.
That’s not magic. It’s just showing up, tediously, for years. Which is exactly why nobody wants to hear it, and exactly why “leaked living brain” outperforms “consistent person kept good notes” by four orders of magnitude.
Same lesson runs through the real story too. Anthropic didn’t need a mythical brain. It needed a .npmignore file and a check in the release pipeline. The unglamorous fix, applied consistently, is nearly always the whole answer. Even the DMCA cleanup afterward, which reportedly took down thousands of unrelated GitHub repositories in the sweep, is a lesson in the same key: the panicked heroic response causes more damage than the quiet preventive one would have.
So, three things worth keeping:
Complexity is not competence. A beautiful graph proves someone made a graph. Ask what it’s for and who maintains it, and watch how quickly the beauty stops mattering.
Check the layer beneath the one you’re correcting. Being right about the mechanism is worth very little if the event never happened. The most confident people in the thread were correcting each other’s descriptions of nothing.
Systems over stories. Boris Cherny’s “fix the process, not the person” is the actual takeaway of this incident, and it’s the one that got the least engagement, which tells you everything about why your feed looks the way it does.
The internet spent a week arguing about whether a company’s second brain had escaped. It hadn’t. There was no brain. There was a missing line in a config file, a person who wasn’t blamed for it, and four strangers who found that story insufficiently exciting and improved it.
Somewhere in there is the most 2026 sentence I can offer you:
The graph wasn’t thinking. Neither, it turns out, were we.
-Hardik
I’ve spent about two decades building data and AI systems, which mostly means I’ve learned to ask “wait, is the thing we’re arguing about real?” roughly one week later than I should have. I write about the gap between what’s actually happening in technology and the story it becomes on the way to your feed. If this made you want to check the layer underneath, subscribe. I’ll keep digging.
https://goelh.substack.com/
https://medium.com/@hardik.goel214
Sources / further reading: CNBC, “Anthropic leaks part of Claude Code’s internal source code” (March 31, 2026) and Anthropic’s statement that this was “a release packaging issue caused by human error, not a security breach,” with no customer data or credentials exposed. InfoWorld / CSO, “Anthropic employee error exposes Claude Code source,” on the .map source-map mechanism and expert commentary (Tanya Janca, Joseph Steinberg). Security researcher Chaofan Shou’s original disclosure post. Lean Blog, “Anthropic’s Claude Code Leak: Why the Instinct to Fire Someone Is the Lazy Response,” documenting Boris Cherny’s public response that nobody was fired and the failure was a process issue. Community Notes and follow-up reporting confirming that the widely shared “I was fired from Anthropic” confession posts were fabricated, and that the individual who claimed responsibility never worked at Anthropic (the post was promotion for his own product). The “Obsidian brain / 21 inputs / ReLU / firing in real time” claim originated in a viral X post and is not supported by any reporting; no such leak has been confirmed. Reports that Anthropic’s subsequent DMCA takedown sweep affected thousands of unrelated GitHub repositories.