Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-105751 Overview
CVE-2026-105751 is a path traversal vulnerability [CWE-22] in Docling, an open-source document processing library that parses diverse formats and integrates with the generative AI ecosystem. The flaw resides in docling/backend/opendocument_backend.py and affects versions from 2.107.0 up to (but not including) 2.120.3. When processing OpenDocument Format (ODF) files, the _image_ref_from_odf_image function reads attacker-controlled xlink:href paths from draw:image elements without URI scheme validation, extraction-directory confinement, or honoring the enable_local_fetch setting. The result is unauthorized local file disclosure through crafted ODF documents.
Critical Impact
An attacker can craft a malicious ODF document that causes Docling to read arbitrary local files from the host filesystem. Readable files decodable by Pillow are embedded in converted output, and other file existence can be probed through read attempts.
Affected Products
- Docling versions 2.107.0 through 2.120.2
- Any application or pipeline using the affected Docling ODF backend
- Generative AI document ingestion workflows relying on vulnerable Docling releases
Discovery Timeline
- 2026-10-05 - CVE-2026-105751 published to NVD
- 2026-10-06 - Last updated in NVD database
Technical Details for CVE-2026-105751
Vulnerability Analysis
Docling's ODF backend parses draw:image elements that reference images via the xlink:href attribute. In normal operation, these references point to entries inside the ODF archive (which is a ZIP container). When the referenced entry is not found inside the archive, the vulnerable code falls back to interpreting the xlink:href value as a filesystem path and attempts to read it directly.
The _image_ref_from_odf_image function performs no scheme validation (so file:// or absolute paths are accepted), no containment check against the extraction directory, and does not consult the enable_local_fetch toggle that other Docling backends use to gate local disk access. Successfully read files that Pillow can decode as images are embedded in the converted output document, directly exfiltrating file contents. Even non-image files produce observable differences based on read success or failure, enabling filesystem reconnaissance.
Root Cause
The root cause is missing input validation on an externally supplied path value combined with an insecure fallback path. The ODF backend trusts the xlink:href attribute without normalizing it, confining it to the extracted archive directory, or enforcing the project's existing enable_local_fetch safety control.
Attack Vector
Exploitation requires an attacker to supply a malicious ODF file (.odt, .ods, .odp) to a Docling-based conversion pipeline. User interaction is required to trigger conversion. Once Docling processes the file, embedded draw:image elements with absolute or relative paths cause the backend to read arbitrary files that the Docling process can access, such as configuration files, credentials, or SSH keys.
# Patch excerpt from docling/backend/utils/image_resource_loader.py
# Commit 6ee9965adfc364b3696a67c825ec2f0d9cee4311
"""Shared image-resource loading for declarative backends.
Turns an image source (a URI, a local file, or a remote URL) into a DoclingDocument
`ImageRef`, enforcing the relevant safety limits: a path-traversal guard when resolving
relative paths, the `enable_local_fetch` / `enable_remote_fetch` toggles, and the
base64 and remote download size caps.
"""
import base64
Source: GitHub Commit 6ee9965. The fix routes ODF xlink:href resolution through the shared ImageResourceLoader, which applies the path-traversal guard and the enable_local_fetch toggle already used by the HTML and Markdown backends.
Detection Methods for CVE-2026-105751
Indicators of Compromise
- ODF files (.odt, .ods, .odp) containing draw:image elements whose xlink:href attributes reference absolute filesystem paths or use ../ traversal sequences
- Converted output documents from Docling that unexpectedly embed images matching the content of system files (for example, SSH keys rendered as text-like image data)
- Docling process read attempts against sensitive paths such as /etc/passwd, /root/.ssh/, or application configuration directories
Detection Strategies
- Inspect inbound ODF archives before conversion by unzipping them and validating that every draw:imagexlink:href points to a relative path contained within the archive
- Enable filesystem auditing (auditd on Linux) on the Docling service account to flag reads outside expected working directories
- Review Docling conversion logs for warnings about missing archive parts followed by image-loading activity
Monitoring Recommendations
- Monitor the Docling service user for unexpected file reads under /etc, /home, /root, or application secret stores
- Alert on Docling pipelines producing output documents whose embedded image sizes or counts diverge from input archive contents
- Track installed Docling versions across build and runtime environments to confirm all instances are on 2.120.3 or later
How to Mitigate CVE-2026-105751
Immediate Actions Required
- Upgrade Docling to version 2.120.3 or later in all environments that process untrusted documents
- Audit any document conversion pipeline that accepted ODF uploads between the release of 2.107.0 and the deployment of the patched version
- Rotate credentials and keys that resided on hosts running vulnerable Docling instances exposed to untrusted input
Patch Information
The fix is included in Docling v2.120.3, delivered via Pull Request #4015 and commit 6ee9965. Full details are available in the GHSA-4xhp-xg4w-8ppm advisory. The patch routes ODF image resolution through the shared ImageResourceLoader, enforcing path-traversal guards and the enable_local_fetch setting.
Workarounds
- If immediate upgrade is not possible, pre-process ODF uploads to strip or rewrite draw:imagexlink:href attributes that do not resolve to entries inside the archive
- Run Docling in a sandboxed environment (container with read-only root filesystem, minimal mounted secrets, or an unprivileged user) to limit what arbitrary file reads can disclose
- Reject ODF inputs at the application boundary until the patched version is deployed
# Upgrade Docling to the patched version
pip install --upgrade 'docling>=2.120.3'
# Verify installed version
python -c "import docling; print(docling.__version__)"
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

