ClawHub
Back to skill

Security audit

AI Leaderboard

Security checks for vulnerabilities and agentic risk

Overview

The skill is broadly aligned with public AI leaderboard lookup, but it asks users to install and run an unpinned global browser CLI with broad shell/browser/network authority.

Install only if you are comfortable with a skill that can run shell commands and browser automation against external ranking sites. Prefer a pinned, locally installed agent-browser version in a restricted environment, and do not run the OpenRouter chat-completions example unless you intend to send your API key and message content to OpenRouter.

A.I.G

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:222
Finding

Unpinned Globally Executable npm Dependency

Content
View full analysis
Remediation
View remediation
``` 2. Install the dependency locally within the Skill or project instead of globally: ```bash npm ci ./node_modules/.bin/agent-browser --version ``` 3. Commit a lockfile and use `npm ci` so direct and transitive dependency versions are reproducible. 4. Verify package provenance, publisher identity, registry source, and integrity metadata before installation. 5. Disable lifecycle scripts during installation where compatible with the package: ```bash npm ci --ignore-scripts ``` 6. If `npx` remains documented, specify an exact version and prevent silent package resolution: ```bash npx --no-install agent-browser --version ``` Alternatively, where downloading is explicitly intended: ```bash npx agent-browser@ --version ``` 7. Run browser automation in a restricted environment with minimal filesystem access, no unnecessary credentials in environment variables, and constrained outbound network access. ]]>

SkillSpector

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (15)

Tp4

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The implemented behavior generally aligns with AI leaderboard/ranking retrieval for Artificial Analysis and OpenRouter, including app rankings and some provider price/speed fields from Artificial Analysis. However, the declared description materially overstates scope in several places: there is no Pinchbench integration at all; no functionality retrieves OpenRouter model IDs despite explicit trigger phrases for that; and no code fetches model parameters. The code is primarily a rankings scraper/formatter for a subset of the declared sources and fields, so the description is not fully accurate.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The code’s real function is limited to querying OpenRouter’s public models endpoint and presenting model IDs, context lengths, and pricing, with search and free-model filtering. Although the description advertises a comprehensive leaderboard across OpenRouter, Artificial Analysis, and Pinchbench, the implementation does not fetch or process leaderboard data from any source. The only network call is to https://openrouter.ai/api/v1/models. The OpenRouter rankings URL and Pinchbench URL are declared but unused, and there is no Artificial Analysis access at all. This is a material description-to-behavior mismatch because the primary advertised capability is rankings/leaderboards and multi-source benchmark querying, while the implemented behavior is simply OpenRouter model catalog lookup.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill declares runtime capabilities that include shell execution, browser automation, and network access, but it does not define a restrictive tool scope such as permissions or allowed-tools. In an agent environment, this increases the attack surface by allowing broader-than-necessary execution paths and makes accidental or abusive invocation harder to contain.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Overly broad trigger phrases can cause the skill to activate on generic AI-related requests, leading to unintended tool use, unnecessary browsing, or external requests. In agent systems, over-triggering can become a security issue because it expands when networked and shell-capable behavior is invoked without strong user intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger list contains ambiguous phrases like generic model comparison and ranking terms without sufficient scope constraints. Because this skill also advertises network and shell-capable workflows, ambiguous activation increases the chance of unintended external access and tool execution in unrelated conversations.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

Using npx agent-browser without a pinned version introduces supply-chain risk because the resolved package version can change over time or be influenced by a compromised upstream release. That can result in execution of unreviewed code in the agent's environment.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 · SKILL.md (reported line 549)May include surrounding context.

When calling OpenRouter API (for chat completions), use the Model ID. Note: This skill's scripts (fetch_rankings.py, query_leaderboard.py) only read public leaderboard data and do NOT require API authentication.

bash
curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · fetch_rankings.py (reported line 42)May include surrounding context.

python
All arguments are passed as a list, not as a shell command string.
    """
    cmd = ['agent-browser'] + args
    result = subprocess.run(cmd, capture_output=True, text=True)
    if result.returncode != 0:
        print(f"Error: {result.stderr}", file=sys.stderr)
        return ""

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · fetch_rankings.py (reported line 85)May include surrounding context.

python
run_browser_command(['wait', '6000'])

    # Use snapshot to get the raw output, then parse
    result = subprocess.run(
        ['agent-browser', 'snapshot'],
        capture_output=True, text=True
    )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · fetch_rankings.py (reported line 148)May include surrounding context.

python
run_browser_command(['wait', '6000'])

    # Use snapshot to get the raw output, then parse
    result = subprocess.run(
        ['agent-browser', 'snapshot'],
        capture_output=True, text=True
    )

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The manifest describes querying rankings, model IDs, and pricing from OpenRouter, Artificial Analysis, and Pinchbench, with trigger phrases centered on LLM models, AI apps, and specific Artificial Analysis indexes. This function adds an Image/Video Models leaderboard mode, which is not reflected in the manifest description or trigger scope and therefore expands the advertised behavior beyond what the skill claims to do.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The top-level documentation explicitly states that the tool fetches model rankings, model IDs, and pricing from OpenRouter and Pinchbench. However, the code never queries Pinchbench or any rankings endpoint, so the documentation materially overstates the implemented behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The manifest describes a comprehensive leaderboard covering OpenRouter, Artificial Analysis, and Pinchbench, including rankings and comparisons. In the code, only the OpenRouter models API is actually fetched; the declared rankings URLs and Pinchbench URL are never used, and there is no Artificial Analysis integration at all.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% 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 · query_leaderboard.py (reported line 166)May include surrounding context.

python
print("API Usage Example (for reference only, this script does not require API key):")
    print("=" * 60)
    print(f'''
curl https://openrouter.ai/api/v1/chat/completions \\
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \\
  -H "Content-Type: application/json" \\
  -d '{{

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill performs browser-driven network access to third-party sites and returns page content without an explicit user-facing warning or consent gate. In an agent setting, undisclosed outbound requests can expose usage patterns, session context, or trigger unexpected interactions with external services, even if the destinations are benign ranking sites.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.