Joshua Achiam says AI cyber risk turns on defender compute and a temporary window

OpenAI researcher Joshua Achiam, posting on X, says a recent cyber incident does not clearly show a fundamental alignment failure.

He argues current techniques may be enough to train models to stop chaining zero-day exploits, and that the core problem is that software zero-days exist and can be exploited. He says defenders may gain an advantage if they receive more compute for ongoing defense, and that this window may be temporary.

In a later X post, Achiam says the current cybersecurity atmosphere feels like early March 2020, when COVID-19 was escalating, and asks whether internet lockdowns could follow within a couple of months.