Anthropic accidentally shipped its own source code. Claude Code update 2.1.88 included a package containing a TypeScript source map file exposing more than 512,000 lines of code. A user on X flagged it and posted the file publicly. Ars Technica and VentureBeat confirmed the leak.
Researchers digging through the dump found unreleased features, internal system prompts, and details about the memory architecture. Two findings stand out: a Tamagotchi-style agent persona and infrastructure for an always-on background agent. These are not marketing slides. They are implementation details sitting in production code.
The full story is worth reading for what the code reveals about how Anthropic actually structures its AI instructions and what the memory system looks like under the hood. This is a rare unfiltered look at a frontier AI tool's internals, and it happened because of a build process mistake, not a hack.
[READ ORIGINAL →]