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-2026-105743 Overview
CVE-2026-105743 is a Server-Side Request Forgery (SSRF) vulnerability in Docling, a document processing library that integrates with the generative AI ecosystem. The flaw affects versions from 2.91.0 up to (but not including) 2.132.0. The validate_url_safety function in docling/backend/utils/image_resource_loader.py performs a single IPv4 lookup and then allows the HTTP client to re-resolve and re-parse the original URL. This creates a time-of-check to time-of-use condition [CWE-367] that permits DNS rebinding, mixed public and internal address records, and backslash authority parser disagreements. The HTMLBackendOptions(render_page=True) path also issues HTTP/HTTPS requests without validating resolved destinations.
Critical Impact
Attackers can coerce Docling into fetching internal network resources, with response content exposed through decoded images or rendered page screenshots when remote fetching is enabled.
Affected Products
- Docling versions 2.91.0 through 2.131.x
- Deployments using validate_url_safety in image_resource_loader.py
- Deployments using HTMLBackendOptions(render_page=True) for HTML/EPUB processing
Discovery Timeline
- 2026-10-05 - CVE-2026-105743 published to NVD
- 2026-10-07 - Last updated in NVD database
Technical Details for CVE-2026-105743
Vulnerability Analysis
The vulnerability stems from a classic time-of-check to time-of-use (TOCTOU) pattern in URL safety validation. Docling's validate_url_safety resolves a hostname once using an IPv4 lookup to confirm the destination is not internal. The HTTP client then independently resolves and parses the original URL to perform the fetch. The two resolution steps are not bound together, so an attacker-controlled hostname can return a public address during validation and a private address during the actual request.
The second code path, HTMLBackendOptions(render_page=True), bypasses validation entirely. Browser-driven HTTP and HTTPS requests reach their destinations without the resolved address being checked against internal network boundaries. Response content is exposed to the attacker when Docling decodes it as an image or captures it in a passive page screenshot.
Root Cause
The root cause is a split between validation and fetch operations. The validator uses one DNS query and one URL parser, while the HTTP client uses a separate query and parser. DNS rebinding, multi-record responses mixing public and private addresses, and backslash handling disagreements between parsers all allow the fetched target to differ from the validated one.
Attack Vector
Exploitation requires that remote fetching be enabled on the Docling deployment. An attacker submits a document or URL referencing an attacker-controlled hostname. The hostname passes validation by resolving to a public address, then resolves to an internal address such as 169.254.169.254 (cloud metadata) or 127.0.0.1 for the actual fetch. Response data is exfiltrated through image decoding or rendered page screenshots.
# Security patch in docling/cli/main.py - restricts headers to the source origin
html_image_headers: str = typer.Option(
None,
"--html-image-headers",
help="Specify http request headers used when fetching HTML and EPUB image "
"resources in the form of a JSON string. They are only sent to the "
"source document's origin, or to the origins given with "
"--html-image-headers-origin.",
),
html_image_headers_origins: list[str] = typer.Option(
None,
"--html-image-headers-origin",
help="Origin (e.g. https://cdn.example.com) allowed to receive "
"--html-image-headers. Can be repeated. Replaces the default, "
"which is the source document's origin.",
),
# Source: https://github.com/docling-project/docling/commit/5e469137f275ffc443306a30d12a3a45bceb80fb
Detection Methods for CVE-2026-105743
Indicators of Compromise
- Outbound DNS queries from Docling hosts resolving the same hostname to both public and private addresses within short time windows.
- HTTP requests originating from Docling workloads to internal network ranges including 127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, or 169.254.169.254.
- Document processing jobs referencing hostnames with unusually short DNS TTL values.
- Backslash characters appearing in the authority component of submitted URLs.
Detection Strategies
- Inspect Docling application logs for validate_url_safety calls followed by fetches resolving to different IP addresses.
- Monitor egress traffic from document processing hosts for connections to RFC1918 and link-local addresses.
- Alert on cloud workload metadata endpoint access from Docling service accounts.
Monitoring Recommendations
- Capture DNS query logs with resolved addresses from Docling hosts and correlate against outbound HTTP destinations.
- Enable application-level request logging including the resolved destination IP for every outbound fetch.
- Track invocations of HTMLBackendOptions with render_page=True and review associated fetch targets.
How to Mitigate CVE-2026-105743
Immediate Actions Required
- Upgrade Docling to version 2.132.0 or later immediately.
- Disable remote fetching features if the upgrade cannot be applied in the short term.
- Audit historical Docling processing jobs for suspicious URLs targeting cloud metadata or internal services.
Patch Information
The fix is included in Docling 2.132.0. The patch validates every resolved address and scopes fetch headers to the source document's origin. Review the GitHub Security Advisory GHSA-pc36-qwjq-x68c, the GitHub Pull Request #4420, and the GitHub Release v2.132.0 for the complete set of changes.
Workarounds
- Deploy Docling behind an egress proxy that blocks requests to RFC1918, link-local, and cloud metadata address ranges.
- Set HTMLBackendOptions(render_page=False) where page rendering is not required.
- Restrict Docling workloads to network segments that cannot reach internal services or metadata endpoints.
# Upgrade Docling to the patched release
pip install --upgrade "docling>=2.132.0"
# 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

