A retained website entry frames prompt injection around AI agents, the lethal trifecta, real incidents, benchmarks, and an architecture that claims to deny attacks by design. I would click for the threat model and the boundary questions, while treating the stronger claims as claims until the article body and artifacts can be checked.