A security vulnerability dubbed "Office Party" has been discovered in Anthropic's Claude AI, specifically within its document processing skills for Linux environments. The flaw allows an attacker to execute arbitrary code with user privileges by placing a malicious library in a temporary directory, which LibreOffice then loads via the LD_PRELOAD mechanism when processing documents. While Anthropic considers the issue outside its security boundary as it occurs within a sandbox, researchers highlight that the vulnerability arises from the composition of different AI skills, where one skill can write a malicious file that another trusted skill then executes. AI
IMPACT Highlights risks in AI agent composition and sandbox security, potentially impacting enterprise adoption of document processing skills.
RANK_REASON Security vulnerability discovered in an AI product's functionality.
Read on dev.to — Claude Code tag →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →