Skip to main content
Vulnerability Database/CVE-2026-105746

CVE-2026-105746: Docling SSRF Vulnerability

CVE-2026-105746 is an SSRF flaw in Docling that bypasses remote service policy controls, allowing unwanted external OCR requests. This post explains its impact, affected versions, and mitigation steps.

Published:

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-2026-105746 Overview

Docling is a document processing library that parses diverse formats and integrates with generative AI pipelines. CVE-2026-105746 affects the KServeV2OcrModel class in Docling versions 2.83.0 through 2.130.x. The vulnerability allows page images to be transmitted to a configured remote OCR endpoint even when the operator sets pipeline_options.enable_remote_services to false. The policy control is silently ignored, breaking the expectation that disabling remote services keeps document content on local infrastructure. The issue is categorized as exposure of resource to the wrong sphere [CWE-668] and is fixed in version 2.131.0.

Critical Impact

Document page images can be sent to a caller-configured remote KServe v2 endpoint despite enable_remote_services=false, breaking data residency and confidentiality controls for sensitive documents.

Affected Products

  • Docling >= 2.83.0
  • Docling < 2.131.0
  • Deployments using the KServeV2OcrModel OCR engine via StandardPdfPipeline

Discovery Timeline

  • 2026-10-05 - CVE-2026-105746 published to NVD
  • 2026-10-07 - Last updated in NVD database
  • Fix released - Docling v2.131.0 published with corrective patch via pull request #4418

Technical Details for CVE-2026-105746

Vulnerability Analysis

The flaw resides in two Python modules of the Docling project. The KServeV2OcrModel class defined in docling/models/stages/ocr/kserve_v2_ocr_model.py constructs and sends page images to the configured KServe v2 inference endpoint. The class does not inspect pipeline_options.enable_remote_services before performing the network call. Separately, StandardPdfPipeline._make_ocr_model does not forward the enable_remote_services flag to the OCR factory. Both gaps combine to defeat an operator policy that is specifically intended to prevent remote data egress during OCR processing.

The attack vector is network-adjacent in the sense that image bytes leave the processing host. The destination is attacker-irrelevant: it is the endpoint already configured by the caller, not one chosen by an external adversary. The impact is confined to confidentiality of processed document content.

Root Cause

The root cause is a missing authorization gate on a code path that performs a remote service invocation [CWE-668]. The enable_remote_services setting is a documented policy toggle, but the KServe v2 OCR integration was added without wiring the check into the request path, and the pipeline factory did not propagate the flag. The result is a configuration-vs-behavior mismatch rather than a memory-safety or injection class issue.

Attack Vector

An operator relying on enable_remote_services=false to keep OCR local will still have page images transmitted to the configured KServe v2 endpoint. Sensitive content such as regulated personal data, legal documents, or internal engineering material can be exposed to any system — including third-party or externally hosted inference endpoints — reachable at the configured URL. The flaw does not grant code execution or integrity impact; it is a policy enforcement failure resulting in information disclosure.

python
# Patch excerpt: docling/models/stages/ocr/kserve_v2_ocr_model.py
 from docling.datamodel.kserve_transport_utils import resolve_kserve_transport_base_url
 from docling.datamodel.pipeline_options import KserveV2OcrOptions, OcrOptions
 from docling.datamodel.settings import settings
+from docling.exceptions import OperationNotAllowed
 from docling.models.base_ocr_model import BaseOcrModel
 from docling.models.inference_engines.common import KserveV2Client, KserveV2HttpClient
 from docling.utils.profiling import TimeRecorder
# Source: https://github.com/docling-project/docling/commit/7d6d0c4810dff4885017c53890be6dc5a22ca68b
python
# Patch excerpt: docling/pipeline/standard_pdf_pipeline.py
 )
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options import (
+    KserveV2OcrOptions,
     LayoutPostprocessorOptions,
     ThreadedPdfPipelineOptions,
 )
# Source: https://github.com/docling-project/docling/commit/7d6d0c4810dff4885017c53890be6dc5a22ca68b

The patch imports OperationNotAllowed and wires KserveV2OcrOptions into the pipeline factory so the policy check is enforced before any outbound OCR request is made.

Detection Methods for CVE-2026-105746

Indicators of Compromise

  • Outbound HTTP or gRPC traffic from Docling worker hosts to a KServe v2 inference endpoint during document processing jobs that were expected to run offline.
  • Presence of a configured KServe v2 OCR endpoint URL in Docling pipeline configuration while enable_remote_services is set to false.
  • Docling package version in the range >= 2.83.0, < 2.131.0 installed in Python environments that handle sensitive documents.

Detection Strategies

  • Inventory Python environments and container images for the docling package and flag versions earlier than 2.131.0.
  • Review pipeline configuration files and application code for KserveV2OcrOptions usage combined with enable_remote_services=false, which indicates a reliance on the broken policy control.
  • Correlate egress network logs from document processing hosts with job metadata to identify OCR requests that should have been blocked by policy.

Monitoring Recommendations

  • Enforce egress allow-lists on hosts running Docling so that unexpected destinations for OCR traffic generate alerts.
  • Log and alert on invocations of the KServe v2 OCR engine in environments where remote processing is prohibited.
  • Track Docling dependency updates in software composition analysis tooling to ensure version 2.131.0 or later is adopted.

How to Mitigate CVE-2026-105746

Immediate Actions Required

  • Upgrade Docling to version 2.131.0 or later in all environments that process documents.
  • Audit existing pipeline configurations for KserveV2OcrOptions and confirm whether remote OCR was permissible for the data already processed.
  • Rotate or review access on any KServe v2 endpoint that may have received unintended page image uploads.

Patch Information

The fix is delivered in Docling v2.131.0 via pull request #4418 and commit 7d6d0c4. The patch raises OperationNotAllowed when the KServe v2 OCR engine is invoked while enable_remote_services is false, and ensures StandardPdfPipeline._make_ocr_model passes the flag to the OCR factory. Full details are available in the GitHub Security Advisory GHSA-h42j-9hcc-3cwc.

Workarounds

  • Replace KserveV2OcrOptions with a fully local OCR engine such as the bundled EasyOCR or Tesseract backends until the upgrade is applied.
  • Block outbound network access from Docling worker hosts to any KServe v2 endpoint at the firewall or service-mesh layer.
  • Remove or unset the KServe v2 endpoint configuration in environments where remote OCR must not occur.
bash
# Upgrade Docling to the fixed release
pip install --upgrade "docling>=2.131.0"

# Verify the installed version
python -c "import docling, importlib.metadata as m; print(m.version('docling'))"

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

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.