Claude Code Leak - Source Zip file inside
It wasn't a sophisticated hack. Instead, it was a classic developer oversight. In the v0.2.1-88 update pushed to npm, Anthropic accidentally included Source Map (.map) files.
For the non-coders: source maps are like a blueprint that allows anyone to turn messy, compressed production code back into the original, human-readable TypeScript.
What was inside?
Nearly 512,000 lines of code were exposed, revealing the "secret sauce" behind one of the world's most advanced AI agents. Key discoveries included:
-
KAIROS: A hidden "proactive agent" designed to work in the background without constant user prompts.
-
System Prompts: The internal "rules" and personality traits Anthropic uses to keep Claude helpful and safe.
-
Future Models: Mentions of unreleased models like Opus 4.7 and Sonnet 4.8.
-
Sentiment Filtering: A hilarious "Fucks" chart used to detect when a user is getting frustrated or abusive.
Should we worry?
Anthropic was quick to clarify that no user data or private API keys were leaked. This was a packaging error, not a breach of their servers.
However, the "toothpaste is out of the tube." While Anthropic is busy issuing DMCA takedowns, the code is already being studied by competitors and enthusiasts. It’s a rare, unfiltered look at how a top-tier AI company builds its logic.
The takeaway? Even the smartest AI companies are prone to human error!