Sources
2,000 People Tried to Jailbreak Simon Willison's AI Assistant — Real-World Prompt-Injection Red-Team Results
Fernando Irarrázaval ran an open challenge at hackmyclaw.com inviting anyone to leak secrets held by an OpenClaw test instance via crafted messages, and ~2,000 people attempted it. Simon Willison's writeup turns it into a rare large-scale, public data point on how AI assistants actually hold up under adversarial prompt injection rather than lab benchmarks. This is directly actionable for anyone shipping tool-using agents that hold credentials or secrets.
↳ Follow the thread