Skip to main content
Vulnerability Database/CVE-2026-105749

CVE-2026-105749: Docling Document Processing DOS Vulnerability

CVE-2026-105749 is a denial of service flaw in Docling that allows attackers to cause sustained CPU usage or memory exhaustion through malicious HTML, JATS, or spreadsheet documents with unbounded span attributes. 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-2026-105749 Overview

CVE-2026-105749 is a resource exhaustion vulnerability [CWE-400] in Docling, a Python library that parses diverse document formats for generative AI pipelines. Versions from 2.0.0 up to but not including 2.131.0 accept unbounded rowspan and colspan attribute values in HTML, JATS, OpenDocument spreadsheet, and BoxNote backends. A crafted document can force the parser into sustained CPU use or multi-gigabyte memory allocation. The document_timeout setting does not interrupt the single backend conversion call, so the process cannot self-recover. The issue is fixed in version 2.131.0.

Critical Impact

A small malicious document can exhaust CPU and memory on hosts running Docling, denying service to downstream AI ingestion pipelines.

Affected Products

  • Docling versions 2.0.0 through 2.130.x
  • Affected backends: docling/backend/html_backend.py
  • Affected backends: docling/backend/jats_backend.py and docling/backend/boxnote_backend.py

Discovery Timeline

  • 2026-10-05 - CVE-2026-105749 published to NVD
  • 2026-10-07 - Last updated in NVD database

Technical Details for CVE-2026-105749

Vulnerability Analysis

The vulnerability lives in Docling's table parsing logic across multiple backends. When processing tables, Docling reads the rowspan and colspan attributes directly from the input document and uses those values to drive loop iterations and grid allocation. No upper bound is enforced before the span values are consumed.

An attacker can submit a document containing a single table cell with span values in the millions or billions. The backend then attempts to allocate a grid proportional to the declared span and iterate over every cell. Export through the TableData.grid property materializes the oversized grid, amplifying memory pressure.

The document_timeout configuration offers no protection. The timeout is checked between conversion stages, not inside the single backend call that performs the span expansion. The parser blocks until the host runs out of memory or CPU, impacting any shared ingestion workload.

Root Cause

The root cause is missing input validation on numeric attributes supplied by untrusted documents. The affected backends trusted rowspan and colspan as parsed, with no clamp against the actual table dimensions or a maximum constant.

Attack Vector

Exploitation requires an attacker to deliver a crafted HTML, JATS, ODS, or BoxNote document to a system running Docling. The attack vector is network-reachable in typical deployments where Docling ingests user-uploaded content for retrieval-augmented generation (RAG) or document indexing pipelines. User interaction is required to submit or trigger processing of the document.

python
 from typing_extensions import override
 
 from docling.backend.abstract_backend import DeclarativeDocumentBackend
+from docling.backend.utils.table_spans import (
+    MAX_COLSPAN,
+    MAX_ROWSPAN,
+    clamp_span,
+    table_width,
+)
 from docling.datamodel.base_models import InputFormat
 from docling.datamodel.document import InputDocument
 from docling.exceptions import DocumentLoadError

Source: GitHub Commit c5b4429

The patch introduces a shared table_spans utility module with MAX_COLSPAN, MAX_ROWSPAN, and a clamp_span helper. Each affected backend imports these and clamps parsed spans to the actual table dimensions before allocating grids.

Detection Methods for CVE-2026-105749

Indicators of Compromise

  • Docling worker processes with resident memory consumption exceeding typical baselines, often reaching multiple gigabytes for a single document.
  • CPU saturation on a single Docling conversion thread lasting significantly longer than historical document processing times.
  • Repeated out-of-memory kills or container restarts correlated with document ingestion events.

Detection Strategies

  • Inspect incoming HTML, JATS, ODS, and BoxNote documents for rowspan or colspan attribute values that exceed the actual table row and column counts.
  • Alert on documents where any span attribute exceeds a reasonable threshold such as 1,000.
  • Correlate document ingestion telemetry with host-level resource spikes to identify suspect inputs.

Monitoring Recommendations

  • Instrument the Docling conversion pipeline with per-document timing and memory metrics.
  • Enforce external process-level resource limits (cgroups, container memory caps) since document_timeout does not break the single backend call.
  • Log the source identity and filename for every document submitted to Docling to support post-incident attribution.

How to Mitigate CVE-2026-105749

Immediate Actions Required

  • Upgrade Docling to version 2.131.0 or later across all ingestion workers and build images.
  • Apply strict container memory and CPU limits to any process invoking Docling backends.
  • Validate and reject untrusted documents containing rowspan or colspan values beyond expected bounds before passing them to Docling.

Patch Information

The fix is included in Docling 2.131.0, delivered via GitHub Pull Request #4414 and commit c5b4429. Full details are published in GitHub Security Advisory GHSA-cgc7-9qp3-86m3 and the v2.131.0 release notes.

Workarounds

  • Pre-process documents to strip or clamp rowspan and colspan attributes to the table's actual dimensions.
  • Run Docling in a sandboxed subprocess with a hard wall-clock timeout and memory ceiling enforced by the operating system.
  • Disable the HTML, JATS, OpenDocument spreadsheet, and BoxNote backends if they are not required for your workload.
bash
# Upgrade Docling to the patched release
pip install --upgrade 'docling>=2.131.0'

# Verify the 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

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.