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-2026-105750 Overview
Docling is a document processing library that parses diverse formats and integrates with generative AI pipelines. CVE-2026-105750 affects Docling versions 2.82.0 through 2.118.0. The HTMLBackendOptions(render_page=True) setting permits file:// URLs because HTMLDocumentBackend._get_browser_request_block_reason does not enforce the enable_local_fetch setting or restrict local requests to the source document directory. Crafted HTML can embed arbitrary readable local text files into a browser-rendered page image when Playwright is installed. The flaw is categorized under [CWE-552: Files or Directories Accessible to External Parties].
Critical Impact
Attackers can exfiltrate readable local files from systems processing untrusted HTML documents through Docling with browser rendering enabled.
Affected Products
- Docling 2.82.0 through 2.118.0 (filesystem Path inputs only)
- Deployments with HTMLBackendOptions(render_page=True) enabled
- Environments where Playwright is installed alongside Docling
Discovery Timeline
- 2026-10-05 - CVE-2026-105750 published to NVD
- 2026-10-07 - Last updated in NVD database
Technical Details for CVE-2026-105750
Vulnerability Analysis
The vulnerability resides in the HTML backend's browser-based rendering path. When callers invoke Docling with HTMLBackendOptions(render_page=True) and pass a filesystem Path as input, the backend launches Playwright to render the document into a page image. During rendering, the browser issues subresource requests for assets referenced in the HTML. The request-blocking logic in HTMLDocumentBackend._get_browser_request_block_reason fails to reject file:// URLs, ignoring the enable_local_fetch configuration and omitting any directory confinement check.
An attacker who supplies a crafted HTML file can embed references to local resources such as /etc/passwd or other readable text files. The browser fetches those files during rendering and embeds their content in the resulting page image, enabling disclosure to anyone who receives the Docling output.
Root Cause
The root cause is missing policy enforcement in the request filter. The function responsible for deciding whether to block a browser request does not consult enable_local_fetch, and it does not constrain file:// requests to the source document directory. Stream inputs are not affected because they use an opaque origin, which the browser treats as untrusted for local reads.
Attack Vector
Exploitation requires that a target workflow processes an attacker-controlled HTML file supplied as a filesystem Path, with render_page=True and Playwright installed. The default configuration, command-line interface, docling-serve, and non-rendering backends are not affected. No authentication is required when Docling is exposed through automated ingestion pipelines that accept third-party documents.
from pathlib import Path
from typing import Any, Final, Iterator, Literal, Optional, Union, cast
from urllib.parse import urlparse
+from urllib.request import url2pathname
from docling_core.types.doc import (
BoundingBox,
Source: GitHub commit 1612b887. The patch adds url2pathname handling so the backend can resolve file:// URLs and enforce directory confinement and the enable_local_fetch policy before dispatching browser requests.
Detection Methods for CVE-2026-105750
Indicators of Compromise
- HTML documents containing <img>, <iframe>, <link>, or CSS url() references pointing to file:// URIs or absolute local filesystem paths.
- Rendered Docling output images that contain fragments of system file content such as /etc/passwd, SSH keys, or configuration files.
- Playwright browser process logs showing navigation or subresource fetches to file:/// locations outside the source document directory.
Detection Strategies
- Scan ingested HTML corpora for file:// schemes and absolute paths in resource attributes prior to processing.
- Inspect Docling output artifacts for embedded text matching known-sensitive file signatures when a rendering pipeline is in use.
- Audit application code and configuration for use of HTMLBackendOptions(render_page=True) together with filesystem Path inputs.
Monitoring Recommendations
- Monitor Docling worker hosts for Playwright child-process reads against sensitive paths such as /etc/, /root/, and credential stores.
- Alert on Docling package versions between 2.82.0 and 2.118.0 reported by software inventory tools.
- Log and review the full input path and backend options for every Docling conversion job in centralized logging.
How to Mitigate CVE-2026-105750
Immediate Actions Required
- Upgrade Docling to version 2.118.1 or later across all document processing environments.
- Disable render_page=True on the HTML backend until the upgrade is complete.
- Route untrusted HTML through stream inputs rather than filesystem Path inputs, which use an opaque origin and are not affected.
Patch Information
The fix is released in Docling 2.118.1. See the GitHub Security Advisory GHSA-q43m-vhcp-mhvm, Pull Request #3948, and the v2.118.1 release notes. The patch updates HTMLDocumentBackend._get_browser_request_block_reason to enforce enable_local_fetch and confine file:// requests to the source document directory.
Workarounds
- Set render_page=False in HTMLBackendOptions to use non-rendering backends that are not affected.
- Uninstall Playwright from Docling runtime environments where browser rendering is not required.
- Process untrusted HTML in sandboxed containers with no read access to sensitive filesystem paths and with outbound file:// resolution blocked.
# Upgrade Docling to the patched release
pip install --upgrade "docling>=2.118.1"
# 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

