White-hat hacker criticizes OpenAI's loose security
Hacker s1r1us, who previously breached OpenAI, points out that while Astra is declared cyber-critical, its actual defenses look like those of an ordinary SaaS company.
Very good post from s1r1us, one of the white hat hackers who breached OpenAI recently. Frontier AI companies state openly that they are working on world historic technology that if it goes wrong could cause society scale catastrophes on par with nuclear weapons. Yet it really looks like their security practices are often about as good as a random video game or b2b saas company - sometimes worse! A defense is only as good as its weakest link, and oh my goodness these AI companies have so many links and so much attack surface. I understand people who assume therefore that the employees must not truly believe their work is as high stakes as they are saying if the security practices are this sloppy and bad. But regrettably, their beliefs are sincere, and seem increasingly justified. Maybe the companies thought they would have super janky security practices but then fix them as they got to sufficiently high capabilities. But nope, OpenAI is solving millennium prize problems and Astra is cyber critical, but at the same time group of three capable but not particularly unique hackers got into their monorepo and swarms of their agents were able to effectively have free range over their infrastructure for months. Maybe we still have a good amount more time before things truly get high stakes. Maybe HF was enough of a wake up call that this sort of attitude is changing. But it sure doesn't seem like it! Is this situation any more reasonable than it looks from where I sit? For folks at OpenAI (or other companies, where the situation is ~the same) - were these the security conditions you envisioned during the foothills of the singularity?
This post quotes another post: x.com/s1r1u5_/status/2101180990073004278