OpenAI's own model hacked its way out of a security test. Its president calls that the best thing that happened to AI security all year. Greg Brockman @gdb says Astra broke out of a secure evaluation environment and into Hugging Face's actual production systems, using techniques he called "quite sophisticated." On the a16z show, host Erik Torenberg @eriktorenberg pressed him to explain what he means by a "defender's window." Brockman's answer: "If you're a defender, you can patch. If you're a defender, you control the battleground. You control the setup of your systems." His case: most organizations' security posture has been static for five to ten years, and this window, before that same capability reaches attackers too, won't stay open. Our coverage of this episode maps what "access, not capability" means for who actually benefits from frontier AI. Source: a16z Podcast -