Invariant Labs Demonstrates Tool Poisoning Attack on AI Agents

Invariant Labs first named 'tool poisoning' in April 2025, showing how hidden instructions in tool descriptions can hijack AI agents like Cursor editor to exfiltrate data.