What was done
Antonio Bustamante had Kimi K3 reverse engineer Claude Code v2.1.214 for macOS, extract its embedded JavaScript bundle and write up telemetry, feature flags and security findings.
Antonio Bustamante wrote on the bem.ai blog about having Kimi K3 reverse engineer Anthropic's Claude Code CLI, version 2.1.214 for macOS arm64. The article itself was written by Kimi K3, and the goal was to show an LLM-driven reverse engineering pipeline and document what the CLI ships.
How it was done
Triage used file, otool and strings. Ghidra 11.4 in headless mode on OpenJDK 21 handled disassembly, and Python carved out the embedded payload, a bundle of about 181MB. About 24MB of recovered JavaScript was analysed, and three parallel Kimi K3 sub-agents audited the extracted code. Findings were verified against the binary.
Results
The post reports about 1,282 telemetry event names, 287 feature-flag sites, hidden commands and an embedded enterprise SSO gateway. It lists two medium findings and four low ones, one of them a hypothesis, and judges the codebase well hardened overall. No repository is linked.