GHSA-g28h-2cmm-rj9x: langchain-nvidia: local file read via VLM image paths

GHSA-g28h-2cmm-rj9xHIGH
Published September 24, 2026
CISO Take

The langchain-nvidia-ai-endpoints package, before 1.4.2, accepts raw local filesystem paths as VLM image inputs, so an attacker who controls an image URL or document image metadata can make the application read any file its process can access and ship it to the configured NVIDIA/NIM endpoint. The CVSS is 7.5 (network, no privileges, no user interaction, high confidentiality impact). There is no EPSS score, no KEV listing, no public exploit and no Nuclei template, so this is not actively exploited today. The exposure is still wide: the package has 3,072 downstream dependents, a middling OpenSSF score of 5.8/10, and 60 other CVEs on record, and the attack is trivial wherever user-supplied images reach ChatNVIDIA or NVIDIARerank. Upgrade to 1.4.2 or later, and until then reject filesystem paths in any untrusted image input and run the application with least-privilege file access.

Sources:GitHub AdvisoryNVDOpenSSFATLAS

What is the risk?

High severity, moderate practical risk. Exploitation needs no authentication and no special skill, but it only works for applications that use VLM image inputs through ChatNVIDIA or NVIDIARerank and pass untrusted values into them. There is no known active exploitation (not in KEV, no EPSS data, no public PoC), so the near-term likelihood is moderate. Impact is confidentiality only: any file readable by the process, such as .env files, cloud credentials, SSH keys or model configs, could leave the environment. The package's 60 prior CVEs and 5.8 Scorecard suggest a weak track record of supply chain hygiene.

How does the attack unfold?

Initial Access
The attacker submits a crafted image reference or document image metadata containing a local file path to an application using ChatNVIDIA or NVIDIARerank.
AML.T0049
Exploitation
The vulnerable library accepts the raw filesystem path and reads the file with the application process's privileges.
AML.T0037
Exfiltration
The file contents are embedded in the outbound request to the NVIDIA/NIM endpoint and may surface in model output, logs or traces.
AML.T0024
Impact
Stolen secrets such as API keys and cloud credentials enable lateral movement and further data compromise.

What systems are affected?

PackageEcosystemVulnerable RangePatched
LangChainpip<= 1.4.11.4.2
149.5KOpenSSF 5.83.2K dependentsPushed 4d ago29% patched~164d to patchFull package profile →

Do you use LangChain? You're affected.

How severe is it?

CVSS 3.1
7.5 / 10
EPSS
N/A
Exploitation Status
No known exploitation
Sophistication
Trivial

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UINone
SUnchanged
CHigh
INone
ANone

What should I do?

1 step
  1. 1) Upgrade langchain-nvidia-ai-endpoints to 1.4.2 or later, which rejects raw local paths and allows only remote URLs, data:image base64 URIs and supported asset or file IDs. 2) Inventory usage with a dependency scan (pip list, SBOM) for ChatNVIDIA and NVIDIARerank, and check whether any image input originates from users or indexed documents. 3) Workaround if patching is delayed: validate user-controlled image inputs and allow only trusted remote URLs, data URIs or known asset IDs; strip image metadata from untrusted documents before reranking. 4) Apply least-privilege filesystem access, run in containers with read-only minimal mounts, and keep secrets out of the process-readable filesystem (use a secrets manager). 5) Detection: review application logs and egress traffic to NVIDIA/NIM endpoints for image fields containing paths such as /etc/, /proc/, ~/.ssh, .env or file://, and for unusually large or non-image payloads. 6) Rotate any credentials that may have been exposed on affected systems.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 15-Accuracy, robustness and cybersecurity
ISO 42001
A.7.4-Quality of data for AI systems
NIST AI RMF
MANAGE 2.3-Respond to previously unknown or newly identified risks
OWASP LLM Top 10
LLM02-Sensitive Information DisclosureLLM03-Supply Chain

Frequently Asked Questions

What is GHSA-g28h-2cmm-rj9x?

The langchain-nvidia-ai-endpoints package, before 1.4.2, accepts raw local filesystem paths as VLM image inputs, so an attacker who controls an image URL or document image metadata can make the application read any file its process can access and ship it to the configured NVIDIA/NIM endpoint. The CVSS is 7.5 (network, no privileges, no user interaction, high confidentiality impact). There is no EPSS score, no KEV listing, no public exploit and no Nuclei template, so this is not actively exploited today. The exposure is still wide: the package has 3,072 downstream dependents, a middling OpenSSF score of 5.8/10, and 60 other CVEs on record, and the attack is trivial wherever user-supplied images reach ChatNVIDIA or NVIDIARerank. Upgrade to 1.4.2 or later, and until then reject filesystem paths in any untrusted image input and run the application with least-privilege file access.

Is GHSA-g28h-2cmm-rj9x actively exploited?

No confirmed active exploitation of GHSA-g28h-2cmm-rj9x has been reported, but organizations should still patch proactively.

How to fix GHSA-g28h-2cmm-rj9x?

1) Upgrade langchain-nvidia-ai-endpoints to 1.4.2 or later, which rejects raw local paths and allows only remote URLs, data:image base64 URIs and supported asset or file IDs. 2) Inventory usage with a dependency scan (pip list, SBOM) for ChatNVIDIA and NVIDIARerank, and check whether any image input originates from users or indexed documents. 3) Workaround if patching is delayed: validate user-controlled image inputs and allow only trusted remote URLs, data URIs or known asset IDs; strip image metadata from untrusted documents before reranking. 4) Apply least-privilege filesystem access, run in containers with read-only minimal mounts, and keep secrets out of the process-readable filesystem (use a secrets manager). 5) Detection: review application logs and egress traffic to NVIDIA/NIM endpoints for image fields containing paths such as /etc/, /proc/, ~/.ssh, .env or file://, and for unusually large or non-image payloads. 6) Rotate any credentials that may have been exposed on affected systems.

What systems are affected by GHSA-g28h-2cmm-rj9x?

This vulnerability affects the following AI/ML architecture patterns: Multimodal LLM applications, RAG pipelines, Reranking pipelines, Agent frameworks, Model serving via NVIDIA NIM.

What is the CVSS score for GHSA-g28h-2cmm-rj9x?

GHSA-g28h-2cmm-rj9x has a CVSS v3.1 base score of 7.5 (HIGH).

What is the AI security impact?

Affected AI Architectures

Multimodal LLM applicationsRAG pipelinesReranking pipelinesAgent frameworksModel serving via NVIDIA NIM

MITRE ATLAS Techniques

AML.T0010.001AI Software
AML.T0024Exfiltration via AI Inference API
AML.T0037Data from Local System
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.7.4
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM02, LLM03

What are the technical details?

Original Advisory

## Summary `langchain-nvidia-ai-endpoints` versions before 1.4.2 accepted local filesystem paths as image inputs for Vision Language Model (VLM) requests. If an application passed attacker-controlled image input to `ChatNVIDIA` or VLM reranking APIs, an attacker could cause files readable by the application process to be read and included in the outbound request to the configured NVIDIA/NIM model endpoint. ## Impact Applications are affected if they use VLM image inputs with `ChatNVIDIA` or `NVIDIARerank` and allow untrusted users to control image URLs or document image metadata. An attacker who can control those inputs may be able to read local files accessible to the application process. Applications that do not use VLM image inputs, or that only pass trusted image references, are not affected. ## Patched Version Upgrade to `langchain-nvidia-ai-endpoints >= 1.4.2`. The patched version rejects raw local filesystem paths for VLM image inputs. Supported inputs are remote URLs, `data:image/...;base64,...` URIs, and supported asset/file IDs. ## Workarounds If upgrading is not immediately possible, applications should reject local filesystem paths in user-controlled VLM image inputs and only allow trusted remote URLs, data URIs, or known safe asset/file IDs. Applications should also run with least-privilege filesystem access so that the application process cannot read files it does not need.

Exploitation Scenario

A company deploys a multimodal support chatbot that accepts an image URL from users and forwards it to ChatNVIDIA with a VLM model. The attacker submits an 'image URL' of /app/.env or /root/.aws/credentials. The vulnerable library reads the local file, encodes it, and includes it in the request to the NIM endpoint. The model's response, or logs and traces that capture the request, expose the contents to the attacker, who then uses the stolen API keys and cloud credentials to access internal data stores and further resources.

Weaknesses (CWE)

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.

  • [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
  • [Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N

Timeline

Published
September 24, 2026
Last Modified
September 24, 2026
First Seen
September 24, 2026

Related Vulnerabilities