Back to skill

Security audit

Qwen Qwen3

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed guide for running Qwen and related local Ollama Herd services, with ordinary setup and local API use rather than hidden or destructive behavior.

Install this in an isolated user environment where possible, avoid running setup commands as administrator/root, and review or pin the PyPI packages before use if you need stricter supply-chain control. Audio, prompts, and generated-content requests are sent to the local Ollama Herd service and may be routed across your configured local fleet.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:42
Finding

Unpinned Third-Party Package Installation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 42 and 51
Vulnerability Type: Unpinned third-party dependencies
Risk Level: Medium

Complete Code Snippet

bash
pip install ollama-herd
bash
uv tool install "mlx-qwen3-asr[serve]" --python 3.14

Technical Analysis

The setup instructions install ollama-herd and mlx-qwen3-asr[serve] without exact version constraints, lockfiles, hashes, or signature verification. Consequently, installations can resolve to package versions and transitive dependencies that differ from those reviewed when the skill was published.

The documented package and repository references point to PyPI and GitHub rather than an evidently malicious source. However, the downloaded packages are not included in this project, so their source code and transitive dependency graphs could not be verified during this audit. This is a supply-chain exposure rather than evidence that the current packages are malicious.

Python package installation and subsequent execution can run third-party code under the installing user's account. A compromised publisher account, malicious future release, or compromised transitive dependency could therefore convert these setup commands into a code-execution path.

Attack Path

  1. An attacker compromises an upstream publisher account, package release process, or transitive dependency.
  2. The attacker publishes a malicious release under a package name used by the setup instructions.
  3. A user follows the documented unpinned installation command.
  4. The package manager resolves and downloads the attacker-controlled release.
  5. Malicious code executes during installation or when the installed application is launched.
  6. The payload acts with the privileges and data access of the user running the command.

Impact Assessment

Successful exploitation could provide arbitrary code execution with the installing user's privileges. Th ...[truncated 372 chars]

Remediation
View remediation

Remediation Suggestions

  • Pin every direct dependency to an exact, reviewed version.
  • Publish and use a lockfile that records resolved transitive dependency versions.
  • Require package hashes, such as with pip's --require-hashes, to verify artifact integrity.
  • Review the source, release provenance, and dependency graph before updating pinned versions.
  • Prefer signed releases or provenance attestations where supported.
  • Install packages in a dedicated virtual environment or another isolated, least-privilege environment.
  • Avoid administrator or root installation unless it is strictly necessary.
  • Document a controlled update process that includes security review and regression testing before changing dependency pins.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (7)

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The top-level title and introductory text describe a Qwen-specific skill centered on Qwen chat, coding, and ASR. Later sections explicitly state that other LLM models, image generation, and embeddings are also available on the same fleet, which contradicts the narrower intent conveyed by the documentation header and introduction.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 52)May include surrounding context.

bash
uv tool install "mlx-qwen3-asr[serve]" --python 3.14
curl -X POST http://localhost:11435/dashboard/api/settings \
  -H "Content-Type: application/json" -d '{"transcription": true}'

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 90)May include surrounding context.

Qwen ASR for transcription

bash
curl http://localhost:11435/api/transcribe -F "audio=@meeting.wav"
python

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 98)May include surrounding context.

md
def transcribe(audio_path):
    with open(audio_path, "rb") as f:
        resp = httpx.post(
            "http://localhost:11435/api/transcribe",
            files={"audio": (audio_path, f)},
            timeout=300.0,

Internal Network Request

Medium
Category
Server-Side Request Forgery
Confidence
70% confidence
Finding

Code issues a request to a loopback, link-local, or private-range host. This can reach internal services not meant to be exposed and is a common SSRF pivot.

Content

Scanner excerpt · SKILL.md (reported line 98)May include surrounding context.

md
def transcribe(audio_path):
    with open(audio_path, "rb") as f:
        resp = httpx.post(
            "http://localhost:11435/api/transcribe",
            files={"audio": (audio_path, f)},
            timeout=300.0,

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest description and the main body of the skill present this as a Qwen-focused skill for local LLM inference, code generation, and speech-to-text. However, the 'Also available on this fleet' section instructs users to invoke image generation and embeddings using different models, which expands behavior beyond the stated Qwen-only purpose.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 166)May include surrounding context.

Image generation

bash
curl -o image.png http://localhost:11435/api/generate-image \
  -H "Content-Type: application/json" \
  -d '{"model":"z-image-turbo","prompt":"a sunset","width":1024,"height":1024,"steps":4}'