Back to skill

Security audit

Model Center

Security checks for vulnerabilities and agentic risk

Overview

This skill mostly matches its NVIDIA API purpose, but it has review-worthy privacy, credential-entry, and package-hygiene issues.

Review this before installing. Use a limited NVIDIA API key, prefer setting it through a protected environment variable instead of typing it into the CLI, avoid sending sensitive data in prompts or embedding text, and treat the unrelated appended log content as a package-quality and provenance concern.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:77
Finding

Unpinned Third-Party Dependency Installation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:77
Vulnerability Type: Unpinned dependency and non-reproducible installation
Risk Level: Medium

Vulnerable Code

text
3. Install dependency: `pip install requests`

Technical Analysis

The documented installation command retrieves the latest version of requests and its transitive dependencies available at installation time. It provides no version constraints, lock file, or package integrity hashes.

Although requests is a legitimate package from the expected public registry, this installation process is not reproducible. A future compromised release, compromised transitive dependency, or unexpected dependency-resolution change would be installed automatically. Package installation may execute build or installation logic with the privileges of the user running pip.

No evidence was found that the project intentionally references a typosquatted package, an untrusted package index, or a currently malicious release. The risk arises from inadequate supply-chain controls.

Attack Path

  1. An attacker compromises a future release of requests or one of its resolved transitive dependencies.
  2. A user follows the documented pip install requests instruction after the compromised version becomes available.
  3. pip resolves and downloads the compromised package because no trusted version or hash is specified.
  4. Malicious installation or runtime code executes in the user's Python environment.
  5. The malicious package gains access to data and permissions available to that user, potentially including NVIDIA_API_KEY when the Skill runs.

Impact Assessment

Successful exploitation could execute arbitrary Python code with the installing or runtime user's privileges. Depending on the environment, this may expose local files, environment variables, NVIDIA API credentials, network access, and data processed by the Skill. The code does not request ...[truncated 123 chars]

Remediation
View remediation

Remediation Suggestions

  1. Add a dependency manifest with an explicitly audited requests version.
  2. Generate and commit a lock file that fixes all transitive dependency versions.
  3. Require package hashes for integrity verification, for example:
    bash
    python -m pip install --require-hashes -r requirements.txt
    
  4. Install dependencies from the official package index over HTTPS and avoid untrusted extra indexes.
  5. Use an isolated virtual environment with only the permissions required by the Skill.
  6. Add automated dependency vulnerability and update monitoring so pinned packages can be upgraded deliberately after review.

T09 · Insecure Skill Coding Practices

Warning
Location
model_center.py:458
Finding

NVIDIA API Key Is Collected Through Echoed Terminal Input

Content
View full analysis

Vulnerability Details

File Location: model_center.py:458-463
Vulnerability Type: Plaintext secret exposure in an interactive terminal
Risk Level: Medium

Vulnerable Code

The non-English source prompt is represented below with equivalent Unicode escapes so that the report remains entirely in English:

python
print("\n\u6ce8\u610f: \u9700\u8981\u8bbe\u7f6eNVIDIA_API_KEY\u73af\u5883\u53d8\u91cf")
print("\u6216\u5728\u6b64\u5904\u8f93\u5165API\u5bc6\u94a5:")
api_key = input("API Key (\u76f4\u63a5\u56de\u8f66\u4f7f\u7528\u73af\u5883\u53d8\u91cf): ").strip()

if api_key:
    center = ModelCenter(api_key)

Technical Analysis

Python's input() function enables normal terminal echo. Consequently, every character of the NVIDIA API key is displayed while the user types it. The exposed key may be observed directly, captured by screen sharing or terminal recording software, or retained in logs produced by systems that record interactive sessions.

The key is subsequently held in memory and placed in an HTTPS Bearer authorization header for the fixed NVIDIA endpoint. The audit found no code that writes the key to disk, logs it, or transmits it to an unrelated destination. Sending the credential to NVIDIA is necessary for the declared API functionality; displaying it during entry is not.

Attack Path

  1. The user starts the command-line interface and selects the chat-testing option.
  2. The application asks the user to enter an NVIDIA API key.
  3. The user enters the key through input().
  4. The terminal visibly echoes the complete key.
  5. A nearby observer, screen-sharing participant, terminal recorder, or session-capture system obtains the displayed credential.
  6. The attacker reuses the credential to authenticate directly to NVIDIA API endpoints until the key is revoked or expires.

Impact Assessment

An exposed key may allow unauthorized consumption of the victim's NVIDIA API quota, a ...[truncated 594 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace the echoed input operation with hidden secret entry:
    python
    from getpass import getpass
    
    api_key = getpass("NVIDIA API key: ").strip()
    
  2. Prefer requiring NVIDIA_API_KEY through a protected environment or secret-management facility rather than collecting it interactively.
  3. Never include the key in logs, exceptions, telemetry, command-line arguments, or persisted configuration.
  4. Keep the credential in memory only for as long as required and avoid unnecessary copies.
  5. Document credential rotation and immediate revocation procedures for suspected exposure.
  6. Where supported by NVIDIA, use narrowly scoped credentials and enforce quota or spending limits to reduce the impact of compromise.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (14)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The documented behavior materially overstates or misstates capabilities, including vision support and polished comparison/browsing flows, while undeclared CLI/local I/O behavior is present. Security-relevant misrepresentation can cause users to rely on controls or data flows that do not actually exist, and hidden interaction paths can process user input in unexpected ways.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill declares use of an API key and documents remote API usage, but it does not explicitly declare tool scope/permissions for environment-variable access and network access. This creates a transparency and least-privilege problem: users and hosting systems may not be clearly informed that prompts and credentials will be used for outbound requests.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The documentation describes chat, embeddings, and image-generation API calls to NVIDIA-hosted services but does not warn that user prompts, text, or other inputs are sent off-device. This is a real privacy and data-handling issue because users may disclose sensitive information without informed consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad enough to activate on ordinary discussion about models, APIs, or NVIDIA topics. Over-broad invocation can cause unintended execution, including unnecessary transmission of user prompts to external services or accidental use of configured credentials.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Large portions of the skill's natural-language content, including the trigger section and appended notes, are written only in Chinese while the rest of the skill is in English. This creates a language/locale constraint that is neither justified nor presented as an opt-in choice, which can violate language policy requirements.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The manifest includes unrelated 'B站学习' and merged-study content that does not belong to the skill's stated purpose. Extraneous embedded content is a supply-chain red flag because it suggests contamination of the package, weak change control, or hidden provenance/data-retention practices that users did not consent to.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The merged study logs preserve opaque identifiers and user-associated strings inside the distributed skill file. This indicates possible retention and disclosure of externally collected data, creating privacy, provenance, and potential tracking risks if those identifiers correspond to users, sessions, or collected artifacts.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · model_center.py (reported line 123)May include surrounding context.

python
}
        
        try:
            response = requests.post(
                endpoint,
                headers=self.headers,
                json=payload,

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · model_center.py (reported line 160)May include surrounding context.

python
}
        
        try:
            response = requests.post(
                endpoint,
                headers=self.headers,
                json=payload,

External Transmission

Medium
Category
Data Exfiltration
Confidence
80% 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 · model_center.py (reported line 185)May include surrounding context.

python
}
        
        try:
            response = requests.post(
                endpoint,
                headers=self.headers,
                json=payload,

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This code sends the assembled chat messages, including user content from the higher-level chat interface, to a remote NVIDIA endpoint via HTTP. While the method has a brief docstring, there is no clear user disclosure here that prompts entered into the skill will be transmitted to an external service, which is the kind of privacy-impacting network action this rule covers.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The function posts the user-supplied prompt and negative prompt to a remote image-generation endpoint. The code describes the API call technically, but it does not provide a user-facing warning that potentially sensitive prompt content will leave the local environment and be processed by an external provider.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This method sends arbitrary input text to the remote embeddings endpoint, which may expose sensitive user data if used on confidential content. There is no user-facing disclosure in this function or surrounding interface that embedding requests involve external network transmission.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
93% confidence
Finding

The inline comment at L359 states that the interactive interface is only for CLI use and that callers in the OpenClaw context should directly invoke ModelCenter methods. However, the same CLI flow later solicits an API key from stdin and reinitializes the client with it, which is a stronger interactive credential-handling behavior than the comment's constrained guidance suggests. This is an intent/documentation mismatch rather than a core functional issue.

Content

No source excerpt is available for this finding.