Cloudanix VulnDB

CVE-2025-6211

Medium · 6.5

LlamaIndex vulnerable to data loss through hash collisions in its DocugamiReader class

CVSS

6.5

EPSS

0.4%

Published

Jul 10, 2025

CISA due date

—

Cloudanix analysis

Monitor only SSVC: Track

A flaw in how some AI document tools create unique labels for text sections can cause important information to be silently replaced or lost, leading to wrong answers from the AI.

Exploitation
Other
Attacker needs
Low Privilege
Complexity
Low
Impact
Data Tampering
Component
Library
Class
Cryptographic Weakness
Remediation
Upgrade
Effort
Low
Patch risk
Low

What it actually does

MD5-based chunk ID generation in DocugamiReader allows collisions for structurally different chunks sharing identical text, causing overwrites that corrupt parent-child hierarchies and stored content.

Why it matters to the business

Organizations relying on llama_index for document processing risk silent loss of critical legal or semantic data, leading to incorrect AI outputs, compliance failures, and degraded retrieval quality across dependent applications.

How an attack unfolds

An attacker supplies documents containing repeated text segments in different structural contexts; the MD5 collision causes one chunk to overwrite another during indexing, breaking hierarchies and injecting incomplete or hallucinated content into downstream AI responses.

How to fix it

Upgrade run-llama/llama_index to version 0.12.41 or later; apply the fixes from commits d38310da63adfc39a5f7eb550e214b8eb02b8b45 and 29b2e07e64ed7d302b1cc058185560b28eaa1352.

How to confirm the fix worked

Confirm package version is at least 0.12.41; test with documents containing repeated text in different structures and verify distinct chunk IDs and intact hierarchies are preserved.

Typical attack path

  1. Identify or craft input documents with duplicate text across distinct sections
  2. Ingest documents via DocugamiReader in vulnerable llama_index versions
  3. Observe overwritten chunk IDs and lost parent-child relationships
  4. Query the resulting index to trigger inaccurate or incomplete AI outputs

How to detect it

Review ingestion logs for repeated chunk IDs on distinct documents; inspect stored chunk hierarchies for unexpected overwrites or missing parent-child links; no public exploitation artifacts are known.

What your security team should do

  • Immediately upgrade all llama_index installations to 0.12.41 or newer
  • Audit existing indexes for signs of prior chunk overwrites
  • Monitor EPSS trend and re-evaluate if exploit maturity increases

The record

A vulnerability in the DocugamiReader class of the run-llama/llama_index repository, up to but excluding version 0.12.41, involves the use of MD5 hashing to generate IDs for document chunks. This approach leads to hash collisions when structurally distinct chunks contain identical text, resulting in one chunk overwriting another. This can cause loss of semantically or legally important document content, breakage of parent-child chunk hierarchies, and inaccurate or hallucinated responses in AI outputs. The issue is resolved in version 0.3.1.

Affected packages

Ecosystem Package Introduced Fixed
PyPI llama-index 0 0.12.41
PyPI llama-index-readers-docugami 0 0.3.1
Chainguard nemo 0 2.3.2-r1

Related vulnerabilities