2,000 people tried to jailbreak a real AI assistant, and Simon Willison wrote it up.

Fernando Irarrázaval ran an open challenge inviting anyone to leak secrets from an OpenClaw test instance, and ~2,000 people attempted it. This is a rare large-scale public data point on how tool-using agents actually hold up under adversarial prompt injection, not a lab benchmark. If you ship agents that hold credentials, read it before you decide your input filter is enough.