Paper Research Agent

• Outcome: Review
• Latest audit: May 29, 2026 · 9:25 PM UTC
• Version: 1.0.0

Overview

This is a coherent paper-research skill, but it can silently change the Python environment and launch many parallel agents without strong user-controlled limits.

Install only if you are comfortable with an autonomous workflow that downloads papers, writes local research folders, may install Python packages, and may spawn many analysis agents. Prefer running it in an isolated environment, preinstalling or pinning dependencies yourself, setting a small paper count/concurrency limit, and reviewing generated tasks before agent execution.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Behavioral AST: exec() Call, eval() Call, Dynamic Import
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (7)

subprocess module call

Medium

• Category: Dangerous Code Execution

Content

try:
            __import__(dep)
        except ImportError:
            subprocess.run([sys.executable, '-m', 'pip', 'install', dep, '-q'])

ensure_deps()

• Confidence: 95% confidence
• Finding: The script automatically invokes pip to install packages at runtime without user confirmation. This expands the skill's privileges from paper research into modifying the local Python environment and executing package installation logic from external registries, which is risky in agent contexts and can expose users to supply-chain compromise or unwanted system changes.

Context-Inappropriate Capability

Medium

• Confidence: 88% confidence
• Finding: The skill instructs the agent to launch a local Python script via subprocess, which expands the capability from paper analysis into arbitrary code execution on the host. In this context, the subprocess is parameterized by user-driven input and runs from a local path, so compromise of the script or abuse of the invocation chain could lead to unintended system actions beyond the user’s expected research task.

Context-Inappropriate Capability

Medium

• Confidence: 97% confidence
• Finding: Auto-installing Python packages is not necessary for the core logic of searching and downloading papers during a normal skill invocation, and it silently performs an external side effect. In an autonomous agent setting this is more dangerous because it allows unannounced environment mutation and introduces supply-chain exposure whenever the skill is run.

Missing User Warnings

Medium

• Confidence: 91% confidence
• Finding: The README explicitly promises automatic PDF downloads and creation of a local research_output directory, but does not warn users that invoking the skill will perform network access and write potentially many files to disk. In an autonomous agent setting, undisclosed side effects matter because a simple natural-language trigger could cause unexpected external requests, storage consumption, and persistence of downloaded content without informed user consent.

Missing User Warnings

Medium

• Confidence: 84% confidence
• Finding: The workflow automates PDF downloads and local file creation without warning the user, which undermines informed consent and can create unwanted network traffic, storage usage, and potentially unsafe file handling. In a skill that may activate from broad research prompts, silent downloading makes the behavior more dangerous because users may not realize external content is being fetched and persisted locally.

Missing User Warnings

Medium

• Confidence: 86% confidence
• Finding: The skill automatically spawns multiple parallel sub-agents without disclosing that behavior to the user, which can amplify resource consumption, increase attack surface, and multiply the effect of prompt/data injection from downloaded content. Parallel autonomous execution is materially riskier than a single analysis pass, especially when each agent reads untrusted PDFs and produces downstream outputs.

Missing User Warnings

Medium

• Confidence: 98% confidence
• Finding: Running pip automatically and silently is dangerous because package installation executes trust-sensitive logic and changes the host environment without prior warning or consent. In this skill context, which may run autonomously on user systems, the lack of confirmation makes the behavior more concerning than in a manual setup tool.

VirusTotal

65/65 vendors flagged this skill as clean.

Static analysis

No suspicious patterns detected.