Researchers at AIR Security have found a flaw in Anthropic's official DOCX, PPTX and XLSX skills — the bundled capabilities that let Claude read, write and convert Word, PowerPoint and Excel files. They call it "Office Party".

The mechanism is unusually concrete. A helper script called soffice.py launches LibreOffice to convert documents, for instance rendering a Word file as a PDF. In environments where Unix-domain sockets cannot be opened, it falls back to LD_PRELOAD and loads a shared library named lo_socket_shim.so from a shared temporary directory. Before loading it, the script checks that the file exists — but not who put it there or what it contains.

So anyone who plants a malicious library at that path before the next document operation gets code running inside LibreOffice with the privileges of the user running the agent. The flaw affects Linux environments where Unix sockets are blocked, including some container configurations and Claude Code's sandbox. macOS is unaffected because it does not load the library through LD_PRELOAD.

The researchers demonstrated two routes: a manually placed library, and — more interesting — a second, apparently benign skill that unpacks a supporting file into the same directory. When the document skill then runs, it loads the foreign library automatically. That chaining is the point: individually reviewed skills can compose into an attack chain.

AIR says it found the flaw in May and reported it. Anthropic acknowledged it but considers it outside its security boundary, because the vulnerable code only runs inside a sandbox. As of early November the script still used the same fixed path and still loaded whatever it found there. Anthropic did not immediately respond to a request for comment on whether that will change.

The practical lesson is uncomfortable: skills are software dependencies. Anyone deploying them needs an inventory of what is installed — including the official ones — and a way to patch the whole set when a vulnerability lands.