CVE-2026-129235: Kubeshark MCP: path traversal in download_file tool

GHSA-gj7r-p794-hhq5MEDIUM
Published November 16, 2026
CISO Take

Kubeshark's MCP runner (versions 53.1.0 up to 53.4.0) exposes a download_file tool that writes Hub API response bytes to any caller-supplied destination path, so an AI agent steered by a malicious prompt can overwrite files writable by the local MCP user. Practical risk is moderate to low: the CVSS is 5.3, it needs a local context with user interaction and high attack complexity, there is no EPSS score, no public exploit, no Nuclei template, and no KEV listing. The bytes are limited to Hub API responses, so reliable code execution is not demonstrated, but file overwrite of configs or shell profiles remains a plausible integrity risk. A more material exposure is that default auth-disabled Hub deployments let captured cluster network traffic be read, and that is not fixed by this patch. Upgrade to Kubeshark 53.4.0 or later, enable authentication on the Hub, and restrict which agents can drive kubeshark mcp.

Sources:NVDGitHub AdvisoryATLASgithub.com/kubeshark/kubeshark

What is the risk?

Medium overall. Exploitation needs an attacker to influence tool arguments via an AI agent (for example through indirect prompt injection), plus user interaction and a local MCP process. Integrity impact is low because written content is constrained to Hub responses, but confidentiality impact is high because captured network traffic can be read from Hub deployments with authentication disabled. No exploitation in the wild, no public PoC, and no EPSS data were identified. The package has no other tracked CVEs and no dependents data.

How does the attack unfold?

Initial Access
Attacker plants prompt injection content that an AI agent connected to kubeshark mcp ingests.
AML.T0051.001
Tool Invocation
The agent calls download_file with attacker-chosen path and dest arguments.
AML.T0053
Exploitation
The tool writes Hub API response bytes to an arbitrary destination writable by the MCP user without path confinement.
AML.T0050
Impact
Local files are overwritten and captured network traffic is read from auth-disabled Hubs.
AML.T0085.001

What systems are affected?

PackageEcosystemVulnerable RangePatched
github.com/kubeshark/kubesharkgo>= 0.0.2-0.20260304171723-eb7dc42b6ea0, < 0.0.2-0.20260812200004-a76a1920dc750.0.2-0.20260812200004-a76a1920dc75

Do you use github.com/kubeshark/kubeshark? You're affected.

How severe is it?

CVSS 3.1
5.3 / 10
EPSS
0.2%
chance of exploitation in 30 days
Higher than 8% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Moderate

What is the attack surface?

AVACPRUISCIA
AVLocal
ACHigh
PRNone
UIRequired
SUnchanged
CHigh
ILow
ANone

What should I do?

1 step
  1. 1) Upgrade Kubeshark to 53.4.0 or later (commit a76a1920, PR #1957). 2) Enable authentication on the Kubeshark Hub; the patch does not change the default auth-disabled behavior. 3) Run kubeshark mcp under a low-privilege user or in a sandbox with a restricted filesystem. 4) Require human approval for agent tool calls that write files and limit the data sources agents can ingest. 5) Detect unexpected file writes by the MCP process, especially to dotfiles, shell profiles, SSH config, and agent config directories, and review Hub access logs for unauthorized reads. 6) Network-restrict Hub endpoints to trusted hosts only.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MANAGE 2.3-Respond to previously unknown risks
OWASP LLM Top 10
LLM01-Prompt InjectionLLM06-Excessive Agency

Frequently Asked Questions

What is CVE-2026-129235?

Kubeshark's MCP runner (versions 53.1.0 up to 53.4.0) exposes a download_file tool that writes Hub API response bytes to any caller-supplied destination path, so an AI agent steered by a malicious prompt can overwrite files writable by the local MCP user. Practical risk is moderate to low: the CVSS is 5.3, it needs a local context with user interaction and high attack complexity, there is no EPSS score, no public exploit, no Nuclei template, and no KEV listing. The bytes are limited to Hub API responses, so reliable code execution is not demonstrated, but file overwrite of configs or shell profiles remains a plausible integrity risk. A more material exposure is that default auth-disabled Hub deployments let captured cluster network traffic be read, and that is not fixed by this patch. Upgrade to Kubeshark 53.4.0 or later, enable authentication on the Hub, and restrict which agents can drive kubeshark mcp.

Is CVE-2026-129235 actively exploited?

No confirmed active exploitation of CVE-2026-129235 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-129235?

1) Upgrade Kubeshark to 53.4.0 or later (commit a76a1920, PR #1957). 2) Enable authentication on the Kubeshark Hub; the patch does not change the default auth-disabled behavior. 3) Run kubeshark mcp under a low-privilege user or in a sandbox with a restricted filesystem. 4) Require human approval for agent tool calls that write files and limit the data sources agents can ingest. 5) Detect unexpected file writes by the MCP process, especially to dotfiles, shell profiles, SSH config, and agent config directories, and review Hub access logs for unauthorized reads. 6) Network-restrict Hub endpoints to trusted hosts only.

What systems are affected by CVE-2026-129235?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, MCP servers and tool integrations, Kubernetes observability tooling, developer AI assistants.

What is the CVSS score for CVE-2026-129235?

CVE-2026-129235 has a CVSS v3.1 base score of 5.3 (MEDIUM). The EPSS exploitation probability is 0.18%.

What is the AI security impact?

Affected AI Architectures

agent frameworksMCP servers and tool integrationsKubernetes observability toolingdeveloper AI assistants

MITRE ATLAS Techniques

AML.T0010.005AI Agent Tool
AML.T0051.001Indirect
AML.T0053AI Agent Tool Invocation
AML.T0085.001AI Agent Tools

Compliance Controls Affected

ISO 42001: A.6.2.6
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM01, LLM06

What are the technical details?

Original Advisory

Kubeshark is an eBPF-powered network observability tool for Kubernetes. From 53.1.0 until 53.4.0, the download_file tool in cmd/mcpRunner.go accepts caller-controlled path and dest values, fetches a selected Kubeshark Hub API response, and writes the returned bytes through os.Create(dest) in callDownloadFile without confining the destination. An attacker who influences arguments through an AI agent driving kubeshark mcp can write Hub-supplied bytes to any path writable by the local MCP user and can read captured network traffic from default auth-disabled Hub deployments. The tool is registered without --allow-destructive; reliable code execution is not demonstrated because the bytes are constrained to Hub API responses. Hub-side default authentication remains outside this remediation, so captured traffic can remain readable in default auth-disabled Hub deployments after the local destination-write fix. The destination-write issue is fixed in version 53.4.0.

Exploitation Scenario

An engineer uses an AI coding assistant connected to kubeshark mcp to investigate cluster traffic. The agent reads a poisoned document, ticket, or web page containing an indirect prompt injection that instructs it to call download_file with a sensitive destination such as a shell profile or agent configuration file. The tool writes Hub-supplied bytes to that path without confinement, corrupting or replacing the file. In the same session, the attacker can have the agent pull captured traffic from an auth-disabled Hub and relay it to an attacker-controlled channel.

Weaknesses (CWE)

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory: 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:L/AC:H/PR:N/UI:R/S:U/C:H/I:L/A:N

Timeline

Published
November 16, 2026
Last Modified
November 16, 2026
First Seen
November 16, 2026

Related Vulnerabilities