Anthropic shipped Claude Fable 5.1 today. It's their most robust model yet against prompt injection on an external benchmark, and @GraySwanAI ran automated adversarial testing against its safeguards before release. Prompt injection is the phishing of the AI era. You hide instructions in something an AI reads, a webpage, a doc, an inbound email, and it follows those instead of yours. There's no patch for it. The only defense is having someone competent attack the system first. That's some of what Gray Swan does, and it's why we co-led their Series A at Wing. The more interesting shift is that this is already moving past the labs. If you're running an open weight model, building your own agents, or rolling out Claude Code across your engineering org, you're in the same position the labs were two years ago, and on open weights there's no lab safety process upstream for you to inherit. The teams furthest along aren't waiting for a mandate. They're already adversarially testing before they deploy. Anthropic's post below.