Claude Code leak exposes prompts, agent loop

Big picture: this is less about “someone found a wrapper” and more about the implementation details of a production AI coding agent becoming readable, which is valuable both for cloning and for adversarial prompt/security research.

• –The most interesting part is the orchestration layer: how Claude Code sequences tool calls, handles multi-file edits, and constrains its own behavior.
• –Prompt leakage matters because it can expose the exact guardrails, refusal patterns, and edit policies Anthropic uses in real workflows.
• –If the npm source-map report is accurate, this is a packaging/security failure, not a model leak.
• –The practical fallout is likely a wave of copycat local-agent projects and a short-term bump in security scrutiny around Claude Code’s tool and permission model.