Aditya Soni

On July 25, we hacked OpenAI. Two bugs let us take over ChatGPT/Codex accounts of OpenAI employees (+some unaffiliated users) and reach connected services: Outlook, Slack, GitHub, etc. We proved it with a PR in OpenAI’s internal codebase . It took us <72h. 🧵

i have no idea what happens after the next two years. maybe we go irrelevant, maybe we won’t, those are all speculations that are not going to help you now, and for the next two years, i’m fairly confident that being good at security is not going to become irrelevant. if anything, i think this might be one of the most fun periods ever to be a hacker, because you can suddenly do things and aim high that were simply not possible before. and the important thing is, just like before, you still need to be good at security. that requires putting in the hours, understanding systems deeply, developing intuitions, and being more curious and ambitious than the people around you. llms are going to amplify all of this. they are not magically going to make everyone capable of finding serious vulnerabilities and getting paid. look at ctf. gpt-5.6 didn’t suddenly turn random people into the best teams in the world. the strong teams are still the strong teams. look at bug hunting. why are many of the people who were good before still the ones who are good now and printing money with models? you can see something similar happening in math. models can now produce or help prove increasingly crazy mathematics, but can you understand all new conjectures being solved? someone still has to understand the output, verify it, know whether it is interesting, and decide what to do next. i actually think this creates demand for more mathematician than mathematicians loosing their jobs, altho new kind of mathematicians who will have to expect that they are not the heroes anymore security should be similar. there will be incredible demand for people who can understand what models are doing, distinguish real results from bullshit, and point them toward useful attack surface. patching is one incredible problem, and random people wont be able to magically suggest or approve patches to complex vulnerabilities from llms but i think the ambition of security research also needs to change. you should start thinking bigger. your research objective should increasingly be something like compromise this company, not lame bugs maybe compromising a company requires finding a bug in some obscure open-source dependency. maybe it requires a browser bug, maybe multiple chains you need to think bigger and thats where companies going to value you the point is that you need to operate one level higher, coordinate all of these capabilities toward an ambitious, long-horizon objective. these models are incredibly capable brains in vats. somebody still has to connect the brain to reality. you cannot just tell codex “hack openai” and expect a shell. a serious target might require weeks of understanding its architecture, mapping dependencies, finding weird third-party components, testing hundreds of hypotheses, killing the useless ones, and repeatedly redirecting models toward the few paths that actually look promising. so if you can spend the next two years with the curiosity of a kid trying to understand every layer of the internet, combined with the dopamine-junky brain that eventually wants to break those layers responsibly and get paid for it, i think this is an incredible time. map the whole attack surface of companies applications, browsers, operating systems, kernels, hypervisors, cpus, gpus, networks, cloud infrastructure, iot, data centres , whatever interests you then pick something and go deep which has highest impact