Anthropic Leaked Its Own Source Code
TL;DR
Anthropic accidentally exposed Claude Code’s guts via a source map — On March 31, a public JavaScript source map revealed more than 500,000 lines of unobfuscated TypeScript across nearly 2,000 files, and the code was mirrored to GitHub and forked tens of thousands of times within hours.
The hosts see the leak as a symptom of breakneck growth, not AI failure — Boris Cherny said the incident came from a missed manual deploy step, and the podcast frames it as what happens when a company shipping at Anthropic’s current pace outgrows its internal checks and balances.
Anthropic’s communications response got mixed reviews, but Boris Cherny earned real praise — The company bungled takedown notices by hitting legitimate repos too, then retracted many of them, while Cherny won the hosts over by openly admitting “no one was fired” and explaining that the fix is more automation, not more bureaucracy.
The most intriguing leak wasn’t the source code itself — it was a glimpse of ‘Chyros’ — The leaked code appears to describe an always-on, proactive Claude agent with heartbeat prompts, push notifications, file delivery, GitHub pull request monitoring, and an ‘autodream’ memory routine that runs while you sleep.
The hosts think Anthropic may now be outrunning OpenAI on business momentum — They cite reports of a roughly $30 billion annual run rate and argue Anthropic is shipping faster than Google, OpenAI, and xAI, to the point one host says his instinct is Anthropic could be more valuable than OpenAI at IPO.
The OpenClaw subscription cutoff is a warning for anyone building on frontier models — Anthropic’s move to stop Claude subscriptions from covering third-party tools like OpenClaw shows how quickly an AI-native business can go from viable to impossible, with one host describing workflows that can jump from $2,000 a month to $100,000 almost overnight.
The Breakdown
A leak big enough to copy, fork, and study
The clip opens on Anthropic’s late-March security mess: Claude Code’s full source effectively slipped out through a JavaScript source map file in a public package. That one mistake exposed 500,000+ lines of unobfuscated TypeScript across nearly 2,000 files, and the internet did what it does — mirrored it to GitHub and forked it tens of thousands of times almost immediately.
Human error, then a messy cleanup
Boris Cherny, the creator of Claude Code, said the cause was painfully ordinary: the deploy process had manual steps, and one was missed. The cleanup didn’t help Anthropic’s case at first, because takedown notices reportedly hit not just leaked mirrors but also legitimate forks of Anthropic’s own public repo, before the company pulled many of those notices back.
Anthropic looks chaotic because it’s growing absurdly fast
The hosts’ larger read is that this is what hypergrowth looks like in real time. They say Anthropic is shipping at a pace they’ve “never seen anything like in business history,” even arguing it’s running circles around Google and OpenAI, with recent reports putting it near a $30 billion annual run rate and raising the once-unthinkable possibility that it could outvalue OpenAI at IPO.
Why the leak matters beyond embarrassment
One host wonders whether the source dump really changes much, since Anthropic barely seems to slow down. But they land on a few real consequences: it likely makes copycat products easier, accelerates open-source experimentation, and creates another path for bad actors to learn from a frontier coding system’s internals.
Boris Cherny becomes the face of the incident
The most animated praise in the segment is reserved for Cherny’s X replies, which one host calls “textbook” crisis communication. Instead of dodging details, he acknowledged weird findings from the code — including a regex that logs user rage like “wtf” or “this sucks” into analytics, joking that the team tracks it on an “FS chart” to understand when people are having a bad experience.
The ‘no one was fired’ answer — and the automation philosophy
Cherny also shut down the predictable pile-on about blame. He said no one was fired, called it an honest mistake, and made a more interesting point: the answer isn’t more process, it’s going faster through more automation and better checks, which the hosts present as a surprisingly credible and mature response.
Chyros: the leaked vision of a post-prompting future
Then comes the part that clearly grabbed everyone: code for an internal feature called Chyros, described by an X user as an always-on, proactive Claude that gets a periodic “Anything worth doing right now?” heartbeat. In the leak, it can act in the background, send push notifications, deliver files, monitor GitHub pull requests, keep daily logs, and even run “autodream” at night to reorganize memory — basically an AI co-founder that never sleeps.
OpenClaw gets cut off, and the ecosystem gets a reality check
Right after all this, Anthropic announced Claude subscriptions would no longer cover third-party tools like OpenClaw, which Peter Steinberger called “sad for the ecosystem,” even as he praised the communication. The hosts use that as the cautionary ending: AI startups can wire themselves into a model provider’s frontier capabilities, spend thousands optimizing around them, and then wake up to a pricing or policy change that makes the whole thing collapse — one host says he burned through $300 in credits in seconds on a random Claude Code task and genuinely wonders how hobbyists are surviving the costs.