Fake Claude site delivers PlugX malware via sideloading attack

✍️ OpenClawRadar📅 Published: April 19, 2026🔗 Source
Fake Claude site delivers PlugX malware via sideloading attack
Ad

Attack details

A fake website impersonating Anthropic's Claude serves a trojanized installer that deploys PlugX malware. The domain mimics Claude's official site, and visitors who download the ZIP archive receive a copy of Claude that installs and runs as expected while deploying malware in the background.

Technical execution

The fake site offers a file called Claude-Pro-windows-x64.zip. The ZIP contains an MSI installer that installs to C:\Program Files (x86)\Anthropic\Claude\Cluade\ - note the misspelling "Cluade" as a red flag. The installer places a shortcut Claude AI.lnk on the Desktop pointing to Claude.vbs inside the SquirrelTemp directory.

When executed, the VBScript dropper:

  • Locates and runs the legitimate claude.exe from C:\Program Files (x86)\Anthropic\Claude\Cluade\claude.exe
  • Creates a new shortcut Claude.lnk on the Desktop pointing directly to claude.exe
  • Copies three files from SquirrelTemp to the Windows Startup folder: NOVUpdate.exe, avk.dll, and NOVUpdate.exe.dat
  • Launches NOVUpdate.exe with a hidden window (window style 0)
Ad

Malware deployment

This is a DLL sideloading attack (MITRE T1574.002). NOVUpdate.exe is a legitimately signed G DATA antivirus updater that attempts to load avk.dll from its directory. The attacker substitutes a malicious version of avk.dll that reads and decrypts payload from the accompanying .dat file.

This three-component sideloading triad (signed executable, trojanized DLL, encrypted data file) is characteristic of the PlugX malware family, a remote access Trojan tracked since 2008.

Behavior and infrastructure

Sandbox analysis shows NOVUpdate.exe establishes outbound TCP connections to 8.217.190.58 on port 443 within 22 seconds of execution. The IP falls within an Alibaba Cloud-associated address range (8.217.x.x). The malware also modifies the registry key HKLM\System\CurrentControlSet\Services\Tcpip\Parameters.

The dropper script includes anti-forensic measures: after deploying payload files, it writes a batch file ~del.vbs.bat that waits two seconds, then deletes both the original VBScript and the batch file itself.

📖 Read the full source: HN AI Agents

Ad

👀 See Also

OpenClaw Security Gap Addressed by Agentic Power of Attorney (APOA) Spec
Security

OpenClaw Security Gap Addressed by Agentic Power of Attorney (APOA) Spec

A developer has published an open specification called Agentic Power of Attorney (APOA) to address security concerns in OpenClaw, where agents currently access services like email and calendar with only natural language instructions as guardrails. The spec proposes per-service permissions, time-bounded access, audit trails, revocation, and credential isolation.

OpenClawRadar
CodeWall AI Agent Discovers Critical Vulnerabilities in McKinsey's Lilli Platform
Security

CodeWall AI Agent Discovers Critical Vulnerabilities in McKinsey's Lilli Platform

CodeWall's autonomous offensive AI agent gained full read/write access to McKinsey's internal Lilli AI platform database within 2 hours, exposing 46.5 million chat messages, 728,000 files, and sensitive system configurations through SQL injection and IDOR vulnerabilities.

OpenClawRadar
AgentSeal Security Scan Finds AI Agent Risks in Blender MCP Server
Security

AgentSeal Security Scan Finds AI Agent Risks in Blender MCP Server

AgentSeal scanned the Blender MCP server (17k stars) and identified several security issues relevant to AI agents, including arbitrary Python execution, potential file exfiltration chains, and prompt injection patterns in tool descriptions.

OpenClawRadar
Bitwarden Agent Access SDK integrates with OneCLI for secure credential injection
Security

Bitwarden Agent Access SDK integrates with OneCLI for secure credential injection

Bitwarden's new Agent Access SDK enables AI agents to access credentials from Bitwarden's vault with human approval, while OneCLI acts as a gateway that injects credentials at the network layer without exposing raw values to agents.

OpenClawRadar