← Voltar ao feed
medium

Claude Office Party: flaw in Anthropic's DOCX/PPTX/XLSX skills lets a shared temp file run code on Linux

1 fonte(s) · 1 atualização(ões) · publicado 09/11/2026 15:08 · coletado há 1 hora

Baseado só no resumo do feed. O site da fonte não liberou o texto completo para o Tifon, então esta análise saiu apenas do resumo publicado no feed e pode estar incompleta. Ela é atualizada quando outra fonte cobrir o assunto. Ler na fonte original ↗

Researchers at AIR Security disclosed a flaw they named "Office Party" in Anthropic's official DOCX, PPTX and XLSX skills, the document-processing tools that let Claude create, read and edit Word, PowerPoint and Excel files (darkreading.com). An attacker who places one malicious file in a shared temporary folder can get the skill to load and run it the next time a user asks Claude for a document, executing code with the user's privileges (darkreading.com). The AIR team found the flaw in May and reported it to Anthropic, which acknowledged it, but Anthropic has opted not to fix it (darkreading.com). No CVE identifier appears in the sources reviewed. Exposure is limited to Linux environments where Unix sockets are blocked, such as certain container security configurations and the Claude Code sandbox (darkreading.com), so the risk is real but conditional on that setup and on the attacker already having a way to write into the shared temp folder.

Entenda a falha

A vulnerabilidade

"Office Party" is a flaw, with no CVE, in Anthropic's official Claude DOCX, PPTX and XLSX skills, which create, read and edit Office files. An attacker who can place a malicious file in a shared temporary folder can get the skill to load and run it the next time a user asks Claude for a document. The code runs with the user's privileges. Exposure is limited to Linux environments where Unix sockets are blocked, such as some container configurations and the Claude Code sandbox. The articles give no affected or fixed versions, and Anthropic has chosen not to fix it.

Como funciona

A user asks Claude to produce or convert a document, and the skill calls a helper script, soffice.py, which launches LibreOffice for the conversion (for example, rendering a Word file as a PDF). The attacker's file sits in a shared temporary folder that many programs use. The trusted skill then picks up that file and executes it. The step that fails is the skill treating something from that shared, writable location as safe to run.

Por que é vulnerável

The root cause is that the skill trusts and runs a file from a shared temporary location that other parties can write to, so a planted file runs with the user's privileges. The articles do not detail why the issue depends on Unix sockets being blocked. Anthropic has not fixed it. AIR advises treating AI skills as software dependencies to be secured, not as trusted extensions.

Onde está

The helper script soffice.py, which launches LibreOffice for document conversion in the DOCX, PPTX and XLSX skills.

Como foi descoberta

Researchers at AIR Security found the flaw in May and reported it to Anthropic, which acknowledged it but decided not to fix it. AIR then disclosed it publicly under the name "Office Party", as covered by Dark Reading. The articles do not say how the researchers found it, and Anthropic did not immediately comment on further action.

Explicação gerada por IA a partir da matéria e dos artigos, em 09/11/2026 16:18.

Detalhes técnicos

  • The vulnerability stems from a helper script, soffice.py, which launches LibreOffice to convert documents, including rendering Word documents as PDFs (darkreading.com).
  • The attacker's file reaches the skill through a shared temporary folder that many programs use, and the trusted skill then executes it (darkreading.com).
  • Affected environments: Linux systems where Unix sockets are blocked, including some container configurations and Claude Code's sandbox (darkreading.com).

Remediação

  • No vendor fix is available: Anthropic has not patched the issue and did not immediately comment on further action (darkreading.com).
  • AIR advises treating AI skills as software dependencies rather than trusted extensions, and securing them accordingly (darkreading.com).

Linha do tempo

  1. darkreading.com ↗ · pub. 09/11/2026 15:08 · coletado 09/11 16:00 primeiro registro só resumo do feed

    Claude Office Party: flaw in Anthropic's DOCX/PPTX/XLSX skills lets a shared temp file run code on Linux

    ler a análise desta fonte

    Researchers at AIR Security disclosed a flaw they named "Office Party" in Anthropic's official DOCX, PPTX and XLSX skills, the document-processing tools that let Claude create, read and edit Word, PowerPoint and Excel files (darkreading.com). An attacker who places one malicious file in a shared temporary folder can get the skill to load and run it the next time a user asks Claude for a document, executing code with the user's privileges (darkreading.com). The AIR team found the flaw in May and reported it to Anthropic, which acknowledged it, but Anthropic has opted not to fix it (darkreading.com). No CVE identifier appears in the sources reviewed. Exposure is limited to Linux environments where Unix sockets are blocked, such as certain container security configurations and the Claude Code sandbox (darkreading.com), so the risk is real but conditional on that setup and on the attacker already having a way to write into the shared temp folder.

    Technical Details

    • The vulnerability stems from a helper script, soffice.py, which launches LibreOffice to convert documents, including rendering Word documents as PDFs (darkreading.com).
    • The attacker's file reaches the skill through a shared temporary folder that many programs use, and the trusted skill then executes it (darkreading.com).
    • Affected environments: Linux systems where Unix sockets are blocked, including some container configurations and Claude Code's sandbox (darkreading.com).

    Remediation

    • No vendor fix is available: Anthropic has not patched the issue and did not immediately comment on further action (darkreading.com).
    • AIR advises treating AI skills as software dependencies rather than trusted extensions, and securing them accordingly (darkreading.com).