research-analyst

Warn

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: MEDIUMPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [Indirect Prompt Injection] (MEDIUM): The skill is designed to ingest and synthesize untrusted data from external websites via WebSearch and WebFetch. This creates a vulnerability where malicious hidden instructions in web content could subvert the agent's reasoning or instructions. Mandatory Evidence: 1. Ingestion points: WebFetch and WebSearch in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Read, Grep, Glob (filesystem read), WebFetch, WebSearch (network read). 4. Sanitization: Absent.
  • [Data Exposure] (LOW): The inclusion of Read, Grep, and Glob tools allows the agent to access the local file system. While these are standard tools for research, this capability increases the impact of a potential indirect prompt injection, as an attacker could attempt to trick the agent into disclosing the contents of local files found through these tools.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 15, 2026, 12:11 AM