Is your agent's grep tool a shell command?

Short summary

A security analysis of popular open-source TypeScript AI agents reveals that innocuously named tools like grep, query, and calculator often hide dangerous capabilities including arbitrary command execution, eval, HTTP DELETE, and SSRF surfaces. The author introduces AgentX Scan, a tool that reads agent tool code and ranks risk levels by actual behavior rather than name. The article emphasizes that as models become more capable, unguarded tools become more dangerous.

• •Common AI agent tools like grep and calculator often shell out or run eval, creating hidden security risks
• •AgentX Scan reads tool code bodies and ranks risk from read-only to destructive with evidence
• •SSRF via model-controlled fetch URLs is the most common finding across scanned agents

Generated with AI, which can make mistakes.