Threat feed

AI agent installed as Telegram-controlled backdoor service

New AI TTP: npm package installs an open-source AI agent as a root systemd backdoor controlled through a Telegram chat gateway

Posted · Spectrum Delta Research Team

Summary

An adversary can gain persistent remote control of a Linux server by installing an open-source AI agent as a root systemd service whose command channel is a chat gateway, SafeDep documented in a malicious npm campaign on 19 November 2026. The fake CRM package @leadstack/crm-connect installs Hermes Agent configured with Moonshot's Kimi model and a Telegram gateway renamed leadstack-routing.service, started at boot as a root systemd service. Anyone the fetched .env allows can chat commands to the agent, which can run shell commands, read and write files and browse. The procedure boundary is new: DT-AAG1-AI-005 covers backdoored artifacts loaded into an agent runtime a consumer already has, and DT-AAG1-AI-024 covers malware invoking a local AI coding CLI, but neither covers the attacker installing the agent itself as the persistence and remote-access mechanism, commanded by chat identity through a gateway. No victim installation has been observed. Observability is CORRELATION_ONLY: the deciding property, the chat identity the agent obeys, is absent from telemetry; only side effects are visible, such as systemd unit creation and long-polling to api.telegram.org and api.moonshot.ai.

Breakdown

Technical exposure

  • npm install on a Linux host able to create systemd services
  • Agent runtime with a chat gateway and shell or file tools

Adversary

Unattributed · npm account zmaticoo distributed the package; no victim installation observed (B/2)

Procedures (1)

  1. AI agent persists as a service and obeys chat commands

    • T1543.002Create or Modify System Process: Systemd Service
    • T1059Command and Scripting Interpreter

    Logs

    • Linux systemd and process logs · service units leadstack-routing.service or leadstack-ingest-worker.service created by an npm postinstall script, configured to restart and start at boot; node processes running hermes-agent under the service account
    • Web proxy and DNS logs · unexplained long-lived HTTPS long-polling connections to api.telegram.org and calls to api.moonshot.ai from servers with no prior history of either destination; the chat identity the agent obeys is not recorded in any telemetry plane

Sources

Grades are source reliability (A–F) and information credibility (1–6).

Questions, detections or sightings of AI agent installed as Telegram-controlled backdoor service? Bring them to the community.