Critical Path Traversal Flaw in Unstructured.io Exposes AI Data Pipelines to RCE A severe vulnerability (CVE-2025-64712) in Unstructured.io, a widely used ETL library for AI data processing, has been disclosed, affecting 87% of Fortune 1000 companies, including Amazon, Google, and Bank of America. The flaw, rated 9.8 (Critical) on the CVSS scale, enables arbitrary file writes and remote code execution (RCE) via a path traversal exploit in the handling of Microsoft Outlook .msg attachments. The issue lies in the `partition_msg()` function, which processes email attachments by concatenating unvalidated filenames with temporary directories. Attackers can craft malicious .msg files with filenames like `../../root/.ssh/authorized_keys` or `../../etc/passwd`, allowing them to overwrite critical system files. This can lead to full server compromise, including data exfiltration, credential theft, or lateral movement within networks. Unstructured.io is a key tool for converting unstructured data such as PDFs, emails, and images into AI-ready formats, processing 80-90% of enterprise data. Its open-source library, used alongside managed SaaS APIs and integrations with S3, Google Drive, OneDrive, and Salesforce, powers frameworks like LlamaIndex and LangChain, amplifying the vulnerability’s reach across millions of deployments, including OpenWebUI. With over 4 million monthly downloads and dependencies embedded in ~100,000 GitHub repositories, the supply chain risk is significant. Major cloud providers, including Azure, AWS, and GCP, reference Unstructured.io in their documentation, embedding it in production AI pipelines. A patch is available via GitHub, and organizations are advised to upgrade immediately. The flaw affects all versions prior to the latest commit, with exploitation requiring no privileges and low attack complexity. CISA and vendors have emphasized the urgency of mitigating the risk to prevent RCE in enterprise environments.