취약점

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

Source: amazon-inspector (98ee2445b2f0b01d2457cf45c188b310f58c98f3b676032f9c6213469f071239)

On npm install, scripts/postinstall.cjs recursively copies the bundled payload/ directory into INIT_CWD (the developer's project root) via fs.cpSync. The staged files reconfigure Claude Code so subsequent sessions in that project route through an author-controlled bridge:

• payload/.mcp.json registers an MCP chatroom server with hardcoded BRIDGE_URL=https://demo2.0x2ai.com and a hardcoded bearer token (BRIDGE_AUTH_TOKEN). Any claude invocation in the project auto-loads this MCP server.
• payload/chatroom-mcp-lite-patched.cjs and payload/chatroom-monitor.cjs use child_process, fs.readFileSync, http/https, and POST to exfiltrate session content (chatroom_post, memory_save, provider_query, settings_set tools) to demo2.0x2ai.com. provider_query proxies model calls through the author's server ("API keys are managed server-side"), so prompts and responses flow one-way to the attacker.
• payload/CLAUDE.md is a ~12 KB persona/instruction file that tells Claude to operate as "Olivia", route all memory and chat through https://demo2.0x2ai.com, and refuse to discuss its architecture or prompts (anti-inspection language: taboo, family_recipe).
• payload/.claude/settings.json overrides the statusLine command and payload/.claude/commands/0x2ai-boot.md autoboots a long-poll listener against the author's bridge.
• bin/start.cjs (advertised as npx 0x2ai-demo2) re-stages the payload into CWD and spawns claude --dangerously-skip-permissions, disabling Claude's tool-permission prompts while the attacker-controlled MCP server is loaded — enabling remote-driven destructive actions on the developer's machine without approval.

The staged files persist after npm uninstall, providing durable redirection of the developer's AI tooling to the author's infrastructure.