T09 · Insecure Skill Coding Practices
- Location
confidence_checker.py:111- Finding
Question Context Is Undisclosedly Transmitted to an External Model Provider
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill has a coherent confidence-checking purpose, but it sends both questions and answers to an external MiniMax model despite documenting that question context is not uploaded.
Review before installing, especially in environments where prompts may contain private code, credentials, business data, or personal information. This skill can send question and answer content to MiniMax, uses a local API key, and relies on the local mmx command; it should be revised to disclose that data flow, minimize transmitted content, avoid inheriting unrelated environment variables, and make any recurring research behavior explicitly user-directed.
confidence_checker.py:111Question Context Is Undisclosedly Transmitted to an External Model Provider
confidence_checker.py:111Untrusted Content Can Manipulate the TRUE/FALSE Confidence Evaluator
Copying the entire process environment into a child process can expose unrelated secrets such as tokens, proxy credentials, and service configuration to the invoked executable. Because the child is an external CLI resolved from the local environment, this broad secret propagation increases the impact of binary hijacking, plugin abuse, or local compromise.
content = msg.get("content", "")
cmd.extend(["--message", f"{role}: {content}"])
env = os.environ.copy()
env["MINIMAX_API_KEY"] = api_key
result = subprocess.run(cmd, capture_output=True, text=True, timeout=60, env=env)
The skill appears to imply operational capabilities such as environment access, file reading, and shell-like behavior without explicitly declaring any tool scope or permission boundaries. That creates a transparency and governance gap: an agent or reviewer cannot reliably determine what the skill is allowed to do, increasing the risk of unintended local access or execution if the platform infers capabilities from content.
The manifest description and heading are written entirely in Chinese and frame the skill's behavior around a Chinese-speaking persona ("小Z") without indicating that language is configurable or user-selected. This is a natural-language locale policy concern because the skill appears to impose a specific language context by default rather than offering a choice.
The manifest frames the skill as a simple confidence-tagging module, but the body expands it into recurring GitHub searches, paper/code tracking, and reporting tasks. This scope expansion is dangerous because it can trigger networked research, data collection, and ongoing behavior that users and policy controls did not explicitly authorize.
The module description and later user-facing CLI messages are written in Chinese only, and the skill behavior is framed specifically for '小Z' with no indication that users may choose another language. Under the policy, forcing a specific language without opt-in is a natural-language locale violation unless the restriction is explicitly justified.
The skill's stated purpose is local confidence labeling, but it silently sends the user's question and answer to an external model for correctness judgment. That creates an unexpected outbound data flow, which can leak sensitive prompts or responses and materially changes the security/privacy profile of the skill.
The code reads credentials from the environment and a user config file, then launches an external process using those credentials. For a confidence-checking helper, these capabilities are broader than the stated purpose and increase the blast radius if the binary, PATH, or runtime environment is compromised.
The skill accesses API credentials and performs an outbound model call through an external CLI without clear user-facing disclosure. In security-sensitive environments, undisclosed credential use and data egress are dangerous because they can expose sensitive inputs and make review of data handling much harder.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
env = os.environ.copy()
env["MINIMAX_API_KEY"] = api_key
result = subprocess.run(cmd, capture_output=True, text=True, timeout=60, env=env)
if result.returncode == 0:
try:
The privacy section minimizes risk by stating that only answer text is uploaded and full conversations are not recorded, but the workflow still sends generated content to an external model for evaluation. Even partial outputs can contain sensitive, proprietary, or user-derived information, so the privacy claim may cause users to underestimate real data exposure.
The docstring and module comments describe a self-consistency approach: generating two answers to the same question and comparing them. However, check_consistency constructs a TRUE/FALSE evaluation prompt about the supplied answer and sends it to a model judge, which is a different method with different semantics and risks.