A Claude Code packaging mistake exposed the tool’s internal workings

Anthropic suffered a second, distinct disclosure days after the Mythos publishing-cache story. On 31 March 2026, a Claude Code release included a source-map file that exposed the coding tool’s internal source. Ars Technica, The Verge and TechCrunch reported the incident and the company’s response.

The distinction from Mythos is essential. This was a software packaging problem involving the application around the AI model, not the release of model weights. Treating both events as the same leak would erase how they happened and exaggerate what was available. The two stories are linked here as related incidents, not duplicate accounts of one discovery.

Contemporary reporting credited security researcher Chaofan Shou with calling attention to the package. Code copies and analyses then spread rapidly. We do not link those copies or reproduce implementation details. The public-interest question is the failed release process and the implications of exposing a commercially important tool’s internal design.

Anthropic acknowledged that a release had included internal source code and said no sensitive customer data or credentials had been exposed. It characterized the event as a human packaging error rather than a security breach. The confirmed badge concerns the source exposure it acknowledged, not an assumption that every claim made by people examining the files was accurate.

The Verge described apparent unreleased ideas, including an always-on agent and a virtual pet. Those are development clues, not commitments to ship. It is possible for code to contain experiments, abandoned work or features that change substantially before release. The archive therefore separates the confirmed accidental disclosure from speculation about the roadmap. All links lead to reporting rather than repositories, archives or download instructions, and the original per-story illustration contains no copied source code or confidential interface imagery.