Tag
A blog post reports that after 6,000 attempts by over 2,000 people, no one successfully leaked secrets from an AI assistant (powered by Opus 4.6) via prompt injection, highlighting improved model resistance but cautioning against overconfidence.