In February 2026, developer Fernando Irarrázaval launched a public challenge at hackmyclaw.com. He tasked his AI assistant, Fiu, with protecting a secrets file from digital attackers.
The challenge went viral on Hacker News. More than 2,000 attackers sent over 6,000 emails attempting prompt injection-hiding malicious commands in messages to trick the AI. The attacks were creative, using urgent subjects and multiple languages.
None succeeded in extracting the secrets.
The system runs on OpenClaw, an open-source framework, using Anthropic's Claude Opus 4.6 model. Side effects included a temporary Google account suspension and over $500 in API costs.
In a separate test, famed AI jailbreaker Pliny the Liberator also failed to breach a similar OpenClaw setup using advanced techniques.
Anthropic's data shows a 0% success rate for injections against Opus 4.6 in constrained tests. However, research indicates other models are vulnerable over 79% of the time.