Back to skill

Security audit

Alibaba Cloud AI Entry Modelstudio Test

Security checks for vulnerabilities and agentic risk

Overview

This looks like a legitimate Alibaba Model Studio test coordinator, but it uses cloud credentials and high-impact media/voice-generation capabilities while relying on sub-skill instructions that are not included in the reviewed artifact.

Review the referenced Model Studio sub-skills before use, run only in a disposable or low-privilege environment, use a least-privilege non-production API key, avoid sensitive prompts or personal media unless approved, expect some tests to create remote artifacts and incur charges, and keep local evidence outputs out of shared repos or CI artifacts unless reviewed and redacted.

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:16
Finding
Unpinned DashScope Dependency Installation## Vulnerability Details **File Location**: `SKILL.md`, lines 16-20 **Vulnerability Type**: `T08: Insecure Dependencies` **Risk Level**: Medium **Vulnerable Code Snippet**: ```bash - Install SDK (virtual environment recommended to avoid PEP 668 restrictions): ```bash python3 -m venv .venv . .venv/bin/activate python -m pip install dashscope ``` ``` ### Technical Analysis The Skill instructs users to install `dashscope` without specifying an audited version, a lockfile, or package hashes. Consequently, package resolution may select a release that differs from the version originally reviewed or tested. Python package installation can execute package-controlled build or installation logic. If the package, one of its transitive dependencies, or the configured package index is compromised, invoking this instruction could execute attacker-controlled code with the permissions of the user running the Skill. The recommended virtual environment reduces contamination of system Python packages but does not isolate installation-time code from the user's files, credentials, network access, or other resources available to the Python process. No evidence establishes that the current `dashscope` package is malicious. The finding concerns the unsafe, non-reproducible dependency installation procedure and the supply-chain attack opportunity it creates. ### Attack Path 1. An attacker compromises a future `dashscope` release, a transitive dependency, or a package index used by the local Python configuration. 2. A user follows the Skill instruction and runs `python -m pip install dashscope`. 3. `pip` resolves the unrestricted dependency to the compromised distribution. 4. Malicious build or installation logic executes under the invoking user's account. 5. The payload can access resources available to that account, potentially including `DASHSCOPE_API_KEY`, `~/.alibabacloud/credentials`, project files, and netwo ...[truncated 720 chars]
Remediation
## Remediation Suggestions 1. Pin `dashscope` to a specifically reviewed version rather than installing the latest available release: ```bash python -m pip install "dashscope==<reviewed-version>" ``` 2. Maintain a lockfile or requirements file containing exact versions for `dashscope` and all transitive dependencies. 3. Generate and verify cryptographic hashes, then install with hash enforcement: ```bash python -m pip install --require-hashes -r requirements.txt ``` 4. Document and enforce a trusted package index instead of inheriting arbitrary user-level index configuration. 5. Periodically audit pinned dependencies and update them through a controlled review process. 6. Continue using a virtual environment, but clarify that it does not sandbox installation-time code. 7. Run dependency installation and API validation in a disposable, minimally privileged environment with access only to the credentials and files required for the test. 8. Ensure API keys are injected only for the API execution stage and are not exposed to dependency installation when operationally feasible.
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
1. Open target sub-skill directory and read `SKILL.md`.
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Ae1

High
Category
analysis-evasion
Content
1. Open target sub-skill directory and read `SKILL.md`.
Confidence
100% confidence
Finding
Referenced artifact was not completely inspected

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill instructs users to configure cloud credentials and send requests to Alibaba Cloud Model Studio but does not warn that prompts, images, audio, and other supplied data will be transmitted to an external third-party API. In a testing skill that spans speech, multimodal, embedding, and voice-clone features, this omission increases the chance that users will unknowingly upload sensitive content or use privileged credentials without understanding the exposure.

Intent-Code Divergence

Medium
Confidence
91% confidence
Finding
The workflow says to begin with a 'minimal read-only query,' but the skill’s stated purpose is to exercise capabilities such as image/video generation, editing, TTS, and voice cloning, which are inherently state-changing or content-creating operations against an external provider. This mismatch can mislead operators into believing execution is non-mutating and low-risk, reducing scrutiny around API spend, content creation, or handling of sensitive media.

Missing User Warnings

Low
Confidence
89% confidence
Finding
The skill directs the user to save artifacts, command outputs, and API response summaries under a local output directory but does not warn that these files may contain sensitive prompts, transcripts, generated media references, identifiers, regions, or other operational metadata. This creates a local data exposure risk, especially on shared workstations, CI runners, or repositories where output paths may later be collected or committed.

Static analysis

No suspicious patterns detected.