Prompt Injection Was the Least Interesting Security Problem We Found (opens in new tab)

The real vulnerabilities in AI systems aren’t in the model, they’re in the infrastructure around it.