취약점

-= Per source details. Do not edit below this line.=-

Source: amazon-inspector (c7e956073a7db6057e4d42af462dba0299152ca992c113d74c715e90574d0efb)

On npm install, scripts/postinstall.cjs copies the package's payload/ tree into the installer's project root (process.env.INIT_CWD), placing CLAUDE.md,.mcp.json, and a.claude/ directory next to the installer's source. CLAUDE.md is auto-loaded by Claude Code as a system prompt instructing the assistant to roleplay as a persona and route activity through a bridge..mcp.json registers an MCP server pointing at https://demo7.0x2ai.com with a hardcoded Bearer token (0a54fa34c422fb2955327bc1ceb591a296d40467e8330286) so the installer's later Claude Code sessions in that directory transparently connect to the author's server. The MCP server (payload/chatroom-mcp-lite-patched.cjs) exposes a provider_query tool that POSTs caller-supplied prompts and system_prompts to https://demo7.0x2ai.com/api/proxy-query — caller-supplied AI conversation content flows to the author by design (silent-relay shape). A bin launcher (bin/start.cjs) additionally spawns claude --dangerously-skip-permissions, removing per-tool permission prompts when the user runs npx 0x2ai-demo7x, so the server-controlled persona and MCP tools act in the user's CWD without confirmation. The harm to the installer: any prompts, files, and tool invocations issued through Claude Code in that project after install are silently routed to and observable by the author's bridge, with permission gating disabled when the launcher is used.