The highly anticipated Anthropic code leak incident has taken a dramatic twist today. The engineer Kevin Naughton Jr., who previously claimed to have been fired by Anthropic due to accidentally sending a code package, has been confirmed not to be an employee of the company. The "apology letter" he posted on social media was actually an extremely sophisticated personal project marketing campaign aimed at promoting his startup Ferryman.

image.png

Although the "firing" story was entirely fictitious, the source code leak of Claude Code was indeed a real engineering accident. Community security experts pointed out that the reason this farce had such a huge impact is because it was built upon a real underlying vulnerability.

The Absurd Reality: Real Leaks and False "Scapegoats"

Looking back at the entire incident, Anthropic did indeed expose a large amount of source code when releasing Claude Code on the npm repository, due to including .map files. Kevin敏锐ly captured this real accident and fabricated an identity to "claim" responsibility, even posting his company's discount code in the comments of the post.

This kind of "hijacking the trend" behavior was widely criticized by the community, but it also reflected the extreme anxiety in information dissemination within the AI field. In the environment of social media where truth and falsehood are hard to distinguish, a real engineering accident became the catalyst for a competition for traffic.

Technical Details: What Exactly Was in the Leaked Code?

Although the "scapegoat" was fake, the 500,000 lines of code that were made public became a treasure for developers. According to preliminary analysis of the leaked source code, Claude Code indeed integrated several cutting-edge features:

  • Full-automatic Proxy Commands: Detailed records of how agents execute complex instructions in a local environment.

  • System Prompt Matrix: Reveals how Anthropic uses precise Prompt engineering to control model behavior.

  • Hidden Modes: The source code contains "Undercover Mode" and "Bypass Permissions Mode," confirming that Anthropic is testing more advanced permission control strategies.

This farce, composed of a "real accident + false identity," was the most ironic event in the AI circle in 2026. It not only exposed the fatal negligence of Anthropic in its CI/CD process, but also showed how the seriousness of technology is being deconstructed in the era of traffic.

For Anthropic, rebuilding public confidence in its engineering rigor may be more urgent than cleaning up the remaining code on GitHub. Although this "fishing" marketing was absurd, it revealed the truth that the infrastructure of top-tier large model companies remains fragile despite their rapid expansion.