Claude Document-Processing Flaw Could Run Attacker Code on Linux

· Original article ↗

Summary

AIR Security reports that Claude’s Office Party flaw can load an attacker-controlled library during document processing on some Linux systems, potentially executing code with the AI agent user’s privileges. Anthropic has not fixed it.

Key points

  • AIR Security found the Office Party flaw in Anthropic’s official DOCX, PPTX, and XLSX skills.
  • When Unix-domain sockets cannot be opened, the soffice.py helper uses LD_PRELOAD to load a shared library.
  • The script checks for lo_socket_shim.so in a shared temporary directory but does not verify its origin or contents.
  • Researchers demonstrated attacks by placing a malicious library in the directory directly or using another AI skill to unpack a supporting file there.
  • The flaw affects certain Linux configurations, including some container security setups and Claude Code’s sandbox; macOS is not affected.
  • AIR Security says Anthropic acknowledged the flaw but considers it outside its security boundary; the issue remains unresolved.
  • AIR recommends maintaining an inventory of AI skills and add-ons and reviewing and updating them like other software dependencies.

Article Details

Vulnerability Types
  • Unverified shared-library loading enabling malicious code execution
Exploitation Status
reported
Exploit Availability
unknown
Patch Status
unavailable
Workarounds
  • Maintain a live inventory of Claude Code skills and add-ons across employees’ agents.
  • Review official skills like other code and update or patch the skill set when vulnerabilities are published.

MITRE ATT&CK

People

Vendors

Products

Industries

Related Articles