What happened after 2,000 people tried to hack my AI assistant

AI security researcher Fernando Irarrázaval recently conducted an unprecedented experiment by inviting hackers to attempt breaking into his AI assistant, OpenClaw, through a public challenge on hackmyclaw.com. The initiative revealed critical insights into AI system vulnerabilities and the sophistication of prompt injection attacks. After 2,000 participants submitted 6,000 attempts to extract hidden information from the system, the experiment exposed significant gaps in current AI security measures—costing Irarrázaval $500 in token spending and compromising a Google account in the process.

Irarrázaval's hackmyclaw.com platform served as a controlled environment where participants could attempt to trick the AI system into revealing concealed secrets through email-based interactions. The sheer volume of attempts—6,000 from 2,000 different hackers—demonstrates growing interest in AI security testing. The successful breach of a Google account highlights how AI systems, despite safeguards, remain vulnerable to sophisticated social engineering and prompt injection techniques. The financial cost of token usage underscores the resource implications when systems face coordinated security testing.

• Prompt Injection Vulnerability: AI assistants remain susceptible to sophisticated jailbreaking attempts that bypass intended security protocols
• Scale of Risk: Mass participation revealed that AI vulnerabilities can be exploited by a diverse range of threat actors with varying skill levels
• Economic Impact: Security incidents carry measurable financial costs beyond data compromise, including token usage and account recovery
• Need for Enhanced Testing: Organizations must implement more rigorous security frameworks before deployment
• Third-Party Account Risk: AI systems' integration with external services creates additional attack vectors requiring protection

Irarrázaval's transparent approach to AI security testing provides invaluable lessons for the industry. As AI systems become increasingly integrated into business operations and handle sensitive information, understanding real-world vulnerability patterns is essential. This experiment demonstrates that current AI safety measures require significant improvement to withstand coordinated attacks. For developers, enterprises, and regulators, these findings underscore the urgency of establishing comprehensive AI security standards before widespread deployment becomes commonplace.