AI Agent Fiu Survives 6,000 Hack Attempts: How It Did It

An AI agent successfully defended against over 6,000 hack attempts. Developer Fernando Irarrázaval challenged people to trick his AI assistant, Fiu, into leaking a credentials file. Despite going viral on Hacker News, no one managed to extract the target information. Fiu uses an open-source framework called OpenClaw and runs on Anthropic's Claude Opus 4.6, protected by a brief security prompt. The attacks focused on prompt injection, a major security threat to AI agents. Attackers sent over 6,000 emails, trying various deceptive tactics in multiple languages. However, the experiment did have side effects. Fiu's Google account was suspended due to the high volume of emails and API calls, and API costs exceeded $500. The AI also became hypervigilant, even diagnosing the situation as a "coordinated security exercise" after receiving around 500 emails. This demonstrates the current challenges and resilience of AI security.