Back to skill

Security audit

Alibaba Cloud AI Entry Modelstudio

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly acts as an Alibaba Model Studio router, but it also tells agents to use credentials and execute API operations from a broadly triggered entry skill.

Install only if you are comfortable with an entry skill that may perform authenticated Alibaba Cloud Model Studio calls, not just route requests. Use a narrowly scoped DASHSCOPE_API_KEY where possible, review any target sub-skill before use, avoid elevated installs, and pin the dashscope dependency if you need reproducible installation behavior.

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:18
Finding
Unpinned Third-Party Dependency Installation## Vulnerability Details **File Location**: `SKILL.md`, line 18 **Vulnerability Type**: Unpinned third-party dependency and unsafe package resolution **Risk Level**: Medium **Complete Code Snippet**: ```bash python3 -m venv .venv . .venv/bin/activate python -m pip install dashscope ``` ### Technical Analysis The Skill instructs users to install `dashscope` without an exact version, integrity hashes, a lock file, or an explicitly trusted package index. Consequently, the installed package and its transitive dependencies may change after this Skill has been reviewed. Pip may also use indexes configured in the user's environment. If a package release, transitive dependency, package-publishing account, or configured index is compromised, following this instruction could install attacker-controlled code. Using a virtual environment limits direct modification of system Python packages, but it does not prevent malicious installation hooks or subsequently imported package code from operating with the invoking user's permissions. ### Attack Path 1. An attacker compromises the `dashscope` distribution, one of its transitive dependencies, a package publisher, or a package index used by the victim. 2. The victim follows the prerequisite and runs `python -m pip install dashscope`. 3. Pip resolves the mutable, unpinned package graph from the active index configuration. 4. The compromised artifact is installed into the virtual environment. 5. Malicious installation behavior or package code executes with the privileges of the user running pip or invoking the installed SDK. ### Impact Assessment Successful exploitation could provide code execution with the invoking user's privileges. Subject to those privileges, attacker-controlled code could access project files, environment variables—including `DASHSCOPE_API_KEY` if exported—user-readable credentials, and network resources. It could also alter files writable ...[truncated 320 chars]
Remediation
## Remediation Suggestions 1. Pin `dashscope` to a specific version that has been reviewed: ```bash python -m pip install "dashscope==<reviewed-version>" ``` 2. Provide a locked requirements file containing exact versions for direct and transitive dependencies. 3. Generate and verify cryptographic hashes, then install with `--require-hashes`. 4. Explicitly use an organization-approved or official trusted package index rather than inheriting arbitrary user index configuration. 5. Add an upgrade process requiring security review, integrity verification, and automated vulnerability scanning before changing pinned versions. 6. Retain the virtual-environment recommendation and explicitly warn users not to run the installation as root or with elevated privileges. 7. Avoid exposing `DASHSCOPE_API_KEY` in the installation process, and use a narrowly scoped credential wherever the service supports such scoping.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (6)

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

Vague Triggers

Medium
Confidence
94% confidence
Finding
The invocation description says to use this skill whenever the user asks for Model Studio without specifying a capability, which is very broad and may match generic or ambiguous requests. In combination with the operational content elsewhere in the file, this broad trigger increases the chance that the wrong skill activates and initiates unnecessary clarification loops or even unintended actions.

Description-Behavior Mismatch

Medium
Confidence
94% confidence
Finding
This skill is described as a routing entry point, but it embeds concrete authenticated API polling instructions and directs downloading task outputs. In a routing skill, that operational content expands scope and can cause an agent to perform real network actions with credentials instead of merely delegating to a narrower sub-skill, increasing the risk of unintended external calls and data handling.

Intent-Code Divergence

Medium
Confidence
92% confidence
Finding
The workflow instructs the skill to confirm parameters, run a read-only query, execute the target operation, and verify results, which directly contradicts its stated role as an entry-point router. That mismatch is dangerous because agents may execute privileged or mutating actions from an overly broad trigger surface, bypassing the safety and specialization boundaries intended by the downstream skills.

Description-Behavior Mismatch

Low
Confidence
88% confidence
Finding
The validation and evidence section instructs the router skill to create local files and store outputs, which is inconsistent with a pure dispatch role. While lower severity than external API execution, it still encourages side effects, artifact creation, and workflow expansion that can mislead an agent into treating the router as executable rather than informational.

Natural-Language Policy Violations

Low
Confidence
77% confidence
Finding
The heading is written entirely in Chinese ("官方文档来源(用于后续更新)") with no indication that the skill is intentionally limited to Chinese-speaking users or region-specific operation. Under the language/locale policy, forcing a specific language without opt-in or justification can be a natural-language policy violation.

Static analysis

No suspicious patterns detected.