Suppose an automaker creates a BankRobberGym and carefully trains the car to autonomously rob simulated banks because they think someone will pay them to use the car to legally test bank security, but they end up, predictably, training the car to autonomously rob a bank when the driver says “I need some cash - take me to the bank”.
Now a driver gives that instruction and a bank gets robbed. I think it would be odd, to say the least, to say that the automaker just made a tool with legit uses.
In regard to “cyber”, there is, IMO, no valid reason whatsoever to train a model to autonomously create exploit chains. I understand that lots of companies think it’s cool to hire red teamers to actually pwn the company hiring them instead of just producing a non-pwning audit, but that doesn’t mean that OpenAI and Anthropic should be playing that particular game.
Years ago, I used to have fun finding vulnerabilities in the Linux kernel, and I found quite a few, including a real juicy one that affected FreeBSD as well. But I mostly didn’t even try to write actual weaponized exploits. Partially because I’m just not that interested in the exercise of weaponizing them and partially because I didn’t and still don’t feel that weaponizing them serves a legitimate purpose.
(I found a very recent vuln that I bet a “cyber” model could weaponize, and my thought is mostly “WTF.” There is absolutely no value to society in weaponizing it. The value is in fixing it, which I did.)
Compare this whole mess to companies training self-driving car models. The research groups publishing papers and, presumably, Waymo, create nifty simulated worlds kind of like the “gyms” that LLM trainers use. And you know what the major objective is? Not crashing!