AI agent survives 6,000 hack attempts in prompt injection test

Over 6,000 emails from 2,000+ attackers failed to extract the target credentials file from developer Fernando Irarrázaval's AI assistant Fiu. The agent, powered by Claude Opus 4.6 on the OpenClaw framework, withstood all attempts but incurred $500+ in API costs and a Google account suspension.

Featured · Fernando Irarrázaval