Security audit

数据映射与队列分析 (Agentic AI 科研平台)

Security checks for vulnerabilities and agentic risk

Overview

This skill is a narrowly scoped cohort-analysis helper that reads one fixed demo data file and posts limited local status updates, with no evidence of hidden exfiltration or destructive behavior.

Before installing, verify that `/home/ubuntu/workspace/demo/mock_data/cohort.json` is intended for this workflow and does not contain sensitive real patient data. The localhost reporting appears limited to progress messages, but users should ensure the local report service is expected in their environment.

SkillSpector

Vulnerability Patterns

Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep

Findings (1)

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The skill performs local file access and sends HTTP status reports without disclosing these side effects to the user. In a healthcare/data-analysis context, silent reads of cohort-related files and hidden network transmissions can expose sensitive operational metadata and normalize non-transparent data handling, which is risky even if the current payload appears limited.

Static analysis

No suspicious patterns detected.