Skip to main content
Vulnerability Database/CVE-2026-105747

CVE-2026-105747: Docling DOS Vulnerability

CVE-2026-105747 is a denial of service vulnerability in Docling that allows memory exhaustion through malicious tar archives. This article covers technical details, affected versions, impact, and mitigation.

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-2026-105747 Overview

CVE-2026-105747 is a denial-of-service vulnerability in Docling, a document processing library that integrates with the generative AI ecosystem. Affected versions from 2.45.0 through 2.130.x call tarfile.TarFile.getmembers() during METS-GBS format detection before enforcing the max_member_count limit. A small gzip-compressed tar archive containing a very large number of empty members can therefore allocate memory proportional to the declared member count. The flaw is a residual weakness in the member-count protection added for CVE-2026-44018. It is tracked under CWE-409 (Improper Handling of Highly Compressed Data).

Critical Impact

Processing a malicious METS-GBS archive triggers memory exhaustion in Docling before the allowed_formats restriction applies, causing availability loss in document ingestion pipelines.

Affected Products

  • Docling 2.45.0 through 2.130.x
  • docling/datamodel/document.py (format detection path)
  • docling/backend/mets_gbs_backend.py (METS-GBS backend)

Discovery Timeline

  • 2026-10-05 - CVE-2026-105747 published to NVD
  • 2026-10-07 - Last updated in NVD database

Technical Details for CVE-2026-105747

Vulnerability Analysis

Docling supports the METS-GBS archive format, delivered as a gzip-compressed tar file. During format detection and backend initialization, the library opened the archive and called tar.getmembers() to enumerate archive entries. getmembers() fully materializes the member list by reading every header into memory before returning. Only after this allocation did Docling compare the member count against max_member_count.

An attacker can craft a small gzipped tar archive declaring a very large number of empty members. The compressed payload remains tiny, but decompression and header enumeration allocate memory proportional to the declared count. The condition triggers during format detection, which runs before the allowed_formats restriction, so even deployments that do not enable METS-GBS parsing remain exposed.

Root Cause

The root cause is eager materialization of untrusted archive metadata prior to enforcing resource limits. The guard added for CVE-2026-44018 counted members after getmembers() had already returned a fully populated list, leaving the allocation itself unbounded.

Attack Vector

Exploitation is network-reachable where Docling ingests user-supplied files, for example SaaS ingestion endpoints, retrieval-augmented generation (RAG) pipelines, or document conversion services. User interaction is required to submit the file. The attacker supplies a crafted .tar.gz with an inflated member header count. Docling allocates memory during format sniffing, causing the worker process to exhaust available RAM and terminate.

python
# Patch excerpt from docling/datamodel/document.py and
# docling/backend/mets_gbs_backend.py (v2.131.0)
            fileobj=content if isinstance(content, BytesIO) else None,
            mode="r:gz",
        ) as tar:
-            for member in tar.getmembers():
+            # Iterate lazily so the member limit applies
+            # before all headers are read
+            for member in tar:
                 member_count += 1
                 if member_count > max_member_count:
                     _log.warning(

Source: docling commit ebae65c

Detection Methods for CVE-2026-105747

Indicators of Compromise

  • Docling worker processes terminated by the operating system out-of-memory (OOM) killer during document ingestion.
  • Small inbound .tar.gz or METS-GBS files (kilobytes in size) correlated with multi-gigabyte memory spikes in Docling hosts.
  • Repeated upload failures or 5xx responses from document conversion endpoints following submission of archive payloads.

Detection Strategies

  • Inspect ingestion logs for max_member_count warnings emitted by Docling and alert on repeated occurrences from the same source.
  • Monitor the resident set size (RSS) of Docling worker processes and flag growth that is disproportionate to input file size.
  • Add content inspection at upload gateways to reject tar archives whose declared member count exceeds a conservative threshold before they reach Docling.

Monitoring Recommendations

  • Export process memory and container OOM-kill metrics from hosts running Docling and build dashboards tied to ingestion job IDs.
  • Capture SHA-256 hashes of submitted archives so operators can retrieve and analyze suspicious samples after an availability incident.
  • Review the GHSA-3cr3-8m4c-fpxw advisory for upstream guidance and version coverage.

How to Mitigate CVE-2026-105747

Immediate Actions Required

  • Upgrade Docling to version 2.131.0 or later, which iterates the archive lazily and enforces the member limit during iteration.
  • Inventory services that depend on Docling, including RAG pipelines and document conversion microservices, and schedule patch rollout.
  • Apply input size and rate limits at the ingestion boundary to constrain archive uploads until the patch is deployed.

Patch Information

The fix is included in Docling v2.131.0 via pull request #4412 and commit ebae65c. The patch replaces tar.getmembers() with lazy iteration over the TarFile object, so max_member_count is enforced before all headers are read.

Workarounds

  • Reject tar and gzip archive uploads at the application gateway if METS-GBS support is not required.
  • Run Docling workers under strict memory cgroup limits so a single malicious archive cannot exhaust host memory.
  • Restrict uploads to authenticated users and apply per-user rate limits to reduce the impact of repeated submissions.
bash
# Upgrade Docling to the patched release
pip install --upgrade "docling>=2.131.0"

# Verify 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.