Security audit

biomarker-investigation-zhcn

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed PatSnap biomarker research helper that uses external life-science data sources and optional web search, with no evidence of hidden persistence, local data access, destructive actions, or exfiltration.

Install only if you are comfortable connecting PatSnap's external MCP service with your API key and allowing the agent to query life-science databases, with occasional web searches when MCP data is insufficient or freshness is needed. Treat medical, clinical, and patent-risk conclusions as research support that should be checked against authoritative sources and qualified experts.

SkillSpector

Vulnerability Patterns

Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (3)

Description-Behavior Mismatch

Medium

Confidence: 94% confidence
Finding: The skill’s public description says it searches academic and patent literature about biomarkers, but the instructions authorize substantially broader collection and analysis across news, deals, companies, drugs, targets, and clinical trial data. This scope expansion can cause users or calling systems to grant the skill more trust and broader data access than they intended, undermining informed consent and least-privilege expectations.

Description-Behavior Mismatch

Medium

Confidence: 96% confidence
Finding: The skill is presented as a biomarker literature/patent search tool, but internally it also directs the agent to assess potential patent infringement risk. Legal-risk assessment is a materially different function from document retrieval and can lead to higher-stakes analysis than users expect, especially in regulated R&D and IP contexts.

Context-Inappropriate Capability

Low

Confidence: 88% confidence
Finding: The skill includes instructions for open-network searching after MCP retrieval, but this behavior is not declared in the skill’s stated purpose or setup expectations. Hidden use of external web sources can introduce privacy, provenance, and data-governance concerns because users may believe results come only from the disclosed PatSnap/MCP sources.

Static analysis

No suspicious patterns detected.