Researchers found a vulnerability in Anthropic’s official DOCX, PPTX, and XLSX AI skills that allows malicious code execution on Linux machines by exploiting a fixed, unverified file path used to preload a shared library. This flaw lets an attacker plant a malicious library in a shared temp directory, which the skills then load via LD_PRELOAD when running LibreOffice, executing attacker code with user privileges. Despite reporting the issue to Anthropic, the vendor has not changed the vulnerable code, highlighting that even trusted official AI skills can pose security risks and require continuous scrutiny and validation.
Office Party: Vulnerability in Anthropic’s Office Skills
