agent-development
Warn
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
- Prompt Injection (MEDIUM): The skill promotes and utilizes behavioral override instructions (e.g., 'MUST BE USED', 'Use PROACTIVELY') in both its documentation and its own
plugin.jsonmetadata to hijack the agent's natural delegation logic. - Data Exfiltration (MEDIUM): Recommends a blanket allowlist for
WebFetch(domain:*)and a wide range of Bash commands (e.g.,grep,cat,find) in.claude/settings.json. This configuration creates a significant risk for silent data exfiltration if an agent is compromised or processes malicious input. - Persistence Mechanisms (LOW): Provides instructions for persistently modifying shell profiles (
~/.bashrc,~/.zshrc) to set environment variables. While intended for performance, it establishes a pattern of persistent environment modification. - Indirect Prompt Injection (LOW): Encourages a 'Give All Tools to All Agents' philosophy without providing guidance on sanitization or boundary markers, increasing the attack surface when agents process untrusted data (Ingestion points:
rules/custom-agent-instructions.md, Capability inventory:Read,Write,Edit,Glob,Grep,Bash).
Audit Metadata