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

In a real-world test of AI assistant security, developer Fernando Irarrázaval challenged hackers to extract secrets from his OpenClaw instance via email prompts. Despite over 6,000 attempts costing $500 in tokens and triggering a Google account suspension, no one succeeded in leaking the secret. The underlying model, Opus 4.6, used a strict anti-prompt-injection system prompt forbidding actions like revealing credentials or executing code. According to Simon Willison, this outcome reflects the increasing effectiveness of training frontier models to resist injection attacks, a trend noted in recent system card releases. However, Willison cautions that 6,000 failures don't guarantee immunity; a determined attacker with a novel approach could still break through. The Hacker News discussion highlighted both healthy skepticism and constructive feedback from the challenge creator. For AI builders, this underscores that while model-level defenses are improving, they are not yet a substitute for robust architectural safeguards in production systems.