"Office Party": researchers say Claude's official Word, PowerPoint and Excel skills load a planted file on Linux. Anthropic reportedly will not fix it
AIR Security says a helper script in Anthropic's DOCX, PPTX and XLSX skills loads a shared library from a fixed temporary path without checking where it came from, so a planted file could run code with the user's own permissions. It affects Linux where Unix sockets are blocked, including Claude Code's sandbox. Not macOS. Anthropic reportedly considers it outside its security boundary. What is known and what to check.
Updated 10 Nov 2026: see what changedHow urgent
Plan this week
Who is affected
People who run Claude with the official DOCX, PPTX or XLSX skills on Linux where Unix-domain sockets cannot be opened, including some container setups and Claude Code's sandbox. AIR Security says macOS is not affected. No CVE, affected versions or fixed version were given, and no source says it has been used in a real attack
What to do
On shared Linux machines or containers, do not let untrusted users or skills write to the shared temporary folder that Claude uses; list the skills and add-ons your agents have installed and review them like any other code; watch for a statement or fix from Anthropic
Published 10 Nov 2026 · Sources last checked 10 Nov 2026No. 73 / 87
Updates, follow-ups and corrections
- Update · 10 Nov 2026. First version of the story, based on one press report (Dark Reading, 9 Nov 2026) quoting the AIR Security report. We have not read the AIR Security report itself. The findings are the researchers'; Anthropic had not answered the press at the time.
We keep the original text. When something changes, when we follow a story, or when we get something wrong, we add a dated note here instead of editing silently.
What happened
Researchers at AIR Security say they found a flaw they call "Office Party" in Anthropic's official DOCX, PPTX and XLSX skills, the tools that let Claude create, read and edit Word, PowerPoint and Excel files. They published their report on 9 November 2026, according to Dark Reading.
How it works, as reported
A helper script, soffice.py, starts LibreOffice to convert documents, for example to render a Word file as a PDF. Where Unix-domain sockets cannot be opened, it falls back to a workaround that loads a shared library through Linux's LD_PRELOAD. It checks whether a file called lo_socket_shim.so already exists in a shared temporary folder, but does not check where the file came from or what it contains. Someone who can place a malicious file there before the next conversion could get their code run inside LibreOffice, with the permissions of the user running the AI agent.
The researchers showed two ways: placing the file by hand, and using a second, harmless-looking skill that unpacks a file into the same folder.
Who is affected
Linux setups where Unix sockets are blocked, which the report says includes some container security configurations and Claude Code's sandboxSandbox A fenced-off area where a program runs with limited access, so if it misbehaves it cannot touch the rest of the system. Tap again to open the glossary. macOS is not affected, according to AIR Security.
What Anthropic said
AIR says it reported the flaw in May and that Anthropic acknowledged it but considers it outside its security boundary, because the vulnerable code only runs inside a sandbox. As of the report, the script still uses the same fixed path. AIR's view is that code running inside the sandbox still has the user's own files and permissions. Anthropic had not answered Dark Reading's request for comment.
What to do
- Know which skills you run. Keep a list of the skills and add-ons your agents have installed, official ones too.
- Treat skills like software dependencies. Review them as you would any code, and update when a fix is published.
- On shared Linux machines or containers, do not let untrusted users, skills or programs write to the shared temporary folder your agent uses. This is our advice, drawn from how the flaw works.
- On macOS, AIR Security says you are not affected.
- Watch for a statement or fix from Anthropic. We found none.
What we could not confirm
We read one press report, not the AIR Security report. We found no CVECVE A public ID number, like CVE-2026-12345, given to one specific security flaw so everyone can talk about the same bug. Tap again to open the glossary, no list of affected versions and no fixed version. No source says the flaw has been used in a real attack: it was shown by the researchers. We have not tested it ourselves.
Watch the short video
The video loads from YouTube (privacy-enhanced mode) only when you press play.
Comments
No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.
We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.
Sources and verification
We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.
Spotted a mistake? Email [email protected] and we will correct it and note the change.