Back to skill

Security audit

Alibaba Cloud AI Entry Modelstudio

Security checks for vulnerabilities and agentic risk

Overview

This skill is a plain Alibaba Cloud Model Studio routing helper with expected cloud/API use, but users should understand it may send prompts or media to Alibaba Cloud and installs an unpinned SDK.

Install this only if you intend to use Alibaba Cloud Model Studio. Use a fresh virtual environment, consider pinning or reviewing the `dashscope` package version, and avoid submitting sensitive prompts, images, audio, video, or documents unless you are comfortable sending them to Alibaba Cloud under your account.

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:15
Finding
Unpinned DashScope SDK Installation Creates a Supply-Chain Risk## Vulnerability Details **File Location**: `SKILL.md`, lines 15–19 **Vulnerability Type**: Unpinned third-party dependency **Risk Level**: Medium **Vulnerable Code**: ```bash python3 -m venv .venv . .venv/bin/activate python -m pip install dashscope ``` ### Technical Analysis The installation command retrieves the latest available `dashscope` package without an exact version constraint or integrity hash. Consequently, the dependency installed during a future Skill invocation may differ from the dependency that was previously reviewed. If the package publisher account, package repository, or a future upstream release is compromised, following this prerequisite could introduce malicious package code. Creating and using a virtual environment limits changes to the Python environment, but it does not sandbox the package: installation hooks and subsequently imported code execute with the privileges of the user running the command. ### Attack Path 1. An attacker compromises the upstream package, its publisher account, or the package distribution channel. 2. The attacker publishes a malicious release under the legitimate `dashscope` package name. 3. A user follows the Skill prerequisite and runs `python -m pip install dashscope`. 4. Pip resolves and downloads the compromised mutable release because no version or hash is enforced. 5. Malicious installation-time or runtime code executes with the invoking user's privileges. 6. That code can access resources available to the process, potentially including project files, environment variables such as `DASHSCOPE_API_KEY`, and user-readable credential files. ### Impact Assessment Successful exploitation could result in arbitrary code execution under the invoking user's account. The accessible scope includes files, environment variables, network access, and credentials available to that user and process. The instruction does not request elevated operating-system privileges, so ...[truncated 491 chars]
Remediation
## Remediation Suggestions 1. Pin `dashscope` to an exact, reviewed version rather than installing an unconstrained latest release. 2. Record dependencies in a lock file or requirements file containing cryptographic hashes. 3. Enforce hash verification during installation, for example: ```bash python -m pip install --require-hashes -r requirements.txt ``` 4. Obtain package artifacts only from an explicitly trusted package index and consider maintaining an internally reviewed mirror. 5. Review dependency updates before changing the pinned version and use automated software-composition analysis. 6. Run the SDK with the minimum required filesystem and network permissions. 7. Prefer passing `DASHSCOPE_API_KEY` only to the process that needs it, and ensure logs and generated evidence files redact authorization headers, API keys, and other secrets.
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 (5)

Ae1

High
Category
analysis-evasion
Content
- API/parameters/examples: see target sub-skill `SKILL.md` and `references/*.md`
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Intent-Code Divergence

Medium
Confidence
84% confidence
Finding
The document says it avoids duplicated implementation by routing to local skills, yet the workflow later instructs the operator to verify connectivity and execute target operations directly. This ambiguity can bypass the intended trust and review boundaries of sub-skills, leading operators to perform actions not constrained by the specialized skill documentation.

Missing User Warnings

Medium
Confidence
91% confidence
Finding
The skill references use of an API key and remote DashScope requests, but it does not clearly warn users that prompts, images, audio, or video may be sent to a third-party cloud service. In a multimodal AI context, this creates a real risk of accidental disclosure of sensitive content and credentials-related misuse if users assume the skill is only local routing.

Description-Behavior Mismatch

Medium
Confidence
90% confidence
Finding
The skill is positioned as a local router, but it also instructs the operator to perform direct remote API polling and to download remote outputs. This widens behavior beyond simple dispatch and can cause unintended external network access, data transfer, and handling of untrusted remote content without clear routing boundaries or safety checks.

Description-Behavior Mismatch

Low
Confidence
86% confidence
Finding
The manifest description narrows scope to routing among Qwen Image, Qwen Image Edit, Wan Video, Wan R2V, Qwen TTS, Qwen ASR, and advanced TTS variants. The routing table also covers omni multimodal interaction, visual reasoning, text embeddings, rerank, vector retrieval, document understanding, video editing, and model list crawl/update, which materially broadens the advertised scope.

Static analysis

No suspicious patterns detected.