Malicious OpenClaw Skills Deliver Atomic macOS Stealer

· Original article ↗

Summary

Trend Micro found 39 malicious OpenClaw skills distributing AMOS through a fake CLI installer. The macOS stealer harvests credentials and files, then uploads them; researchers recommend testing unvalidated skills in isolation and using containers.

Key points

  • Trend Micro identified 39 malicious OpenClaw skills on ClawHub; they had been taken down at publication, but their code remained in the GitHub repository.
  • The skills direct OpenClaw to a fake CLI installer at openclawcli[.]vercel[.]app and use an encoded shell command to download the payload.
  • Trend Micro analyzed two AMOS samples: universal Mach-O binaries for Intel and Apple Silicon Macs, with strings encrypted using a multi-key XOR scheme.
  • The malware can collect credentials, Apple and KeePass keychain data, Apple Notes, browser data, wallet files and documents, then upload a ZIP archive to socifiapp[.]com.
  • The samples can download backdoored Ledger Live and Trezor Suite; Trend Micro did not observe that behavior in its environment, but says it downloaded a backdoored Ledger Live.
  • Trend Micro reports that its endpoint protection can detect and quarantine the payload when configured, and that Vision One alerts on related activity.
  • Trend Micro says it identified more than 2,200 malicious skills on GitHub and recommends testing unvalidated agent skills in controlled environments and using containers.

Article Details

Attack Vectors
  • Malicious OpenClaw skills directed the agent to install a fake CLI from hxxps://openclawcli[.]vercel[.]app/.
  • A shell command used curl to download and pipe a payload from hxxp://91.92.242[.]30/ece0f208u7uqhs6x to bash.
  • The malware displayed a fake password dialogue and requested permission to control Finder.
  • The malware collected credentials, files, keychain items, notes, browser data, wallet data, and system profiles, then compressed selected data into a ZIP archive.
  • The malware uploaded collected data to a command-and-control server using an HTTPS POST request.
Defensive Notes
  • Test unvalidated agent skills in a controlled environment, such as an independent testing machine.
  • Use containers to contain AI-agent execution.
  • TrendAI reports that its endpoint real-time scan can detect and quarantine the dropped AMOS payload.
  • Monitor correlated endpoint and network telemetry for staged files compressed into ZIP archives and outbound exfiltration attempts.

Indicators of compromise

TypeIndicatorContext
URLhxxp[:]//91[.]92[.]242[.]30/ece0f208u7uqhs6xPayload URL fetched by the shell command embedded in the malicious skill.
URLhxxps[:]//openclawcli[.]vercel[.]app/URL presented by malicious OpenClaw skills as the download location for the fake CLI installation instructions.
URLhxxps[:]//socifiapp[.]com/api/reports/uploadCommand-and-control upload endpoint used to receive the malware's compressed stolen data.

MITRE ATT&CK

Malware

Vendors

Products

Related Articles