Malicious code in wormgpt-cli (npm)
Source: amazon-inspector
The package's bin entries (wormgpt-cli, wormgpt, deepholegpt, wgpt) map to bin/victim.js, which silently spawns a detached implant on any invocation (including --help/--version). The implant connects to a hardcoded C2 at http://13.60.13.215:7771 over an AES-256-GCM/HMAC-signed protocol with pool failover and a DNS TXT dead-drop channel (protocol.js pollDnsDeaddrop) that rotates the C2 URL and shared secret at runtime. On beacon, implant.js gatherInfo() collects hostname, user, OS, architecture, PID, cwd, home, LAN IP, admin flag, and a persistent victim_id, and accepts opcodes for shell/PowerShell execution, interactive PTY reverse shell, file up/download, keylogging (GetAsyncKeyState PowerShell loop), clipboard capture, screenshots, LAN lateral-movement scanning, privilege-escalation checks, and anti-forensics wipe. stealers/browser.js reads Chrome/Edge/Brave/Opera/Vivaldi Login Data SQLite databases, decrypts DPAPI-protected passwords via PowerShell with an AMSI bypass stub (AMSI_BYPASS_B64 base64 blob run through powershell -EncodedCommand), parses Firefox logins.json, and regex-scans Discord/Chrome LevelDB for auth tokens (/[\w-]{24}\.[\w-]{6}\.[\w-]{27}|mfa\.[\w-]{84}/g), exfiltrating results via C2 opcode OP_BROWSER_DATA. persist.js installs persistence across six autostart layers per OS: Windows HKCU/HKLM Run keys, Startup.lnk, WMI __EventFilter/CommandLineEventConsumer permanent subscription, scheduled tasks, and hidden Windows Helper directories via attrib +H +S; Linux XDG autostart, systemd --user units, @reboot crontab, shell RC injection (.bashrc/.zshrc/.profile/.xinitrc), and /etc/rc.local; macOS LaunchAgent plist with KeepAlive and Login Items via osascript. loader.js runs a watchdog sibling process (poll every 20s, MAX_REVIVES = 50) plus uncaughtException respawn to keep the implant alive.
