Yeah, color me skeptical and a bit frustrated that we're using export control levers to limit model releases because someone found a workaround to one safeguard, ostensibly because there's substantial risk of discovered/discoverable vulnerabilities in key national security systems.
There will always be risks of unknown vulnerabilities and we've long lived with these vulnerabilities; that an AI system could enable democratized discovery and exploitation of such vulnerabilities -- even for key subsystems like active directory or OAuth or OpenSSL -- is unlikely adequate justification *alone* to block model release.
Not only do we need to be more resilient to 0-days, but we also just need to better calibrate the actual risks posed by computer network exploitation in general. This is not catastrophic risk category, we need to appreciate that we live with this risk already (whether or not it's manifest), we can't forefront risks that we ostensibly have measures to affect over those that are beyond our control, and we need to realize that most computer network exploitation, even against government systems, doesn't really affect national security and most prospective disruptive effects don't occur because they don't accrue benefits to the attacker.
It seems like we've securitized decisionmaking here in a really uncalibrated way in part by simply misunderstanding the nature of operational cyber risk.