The exploit suite is coded by LLMs (and contains bugs)
The fixes are coded by LLMs
The instructions are generated by LLMs
Humans trust this code, and run the software
What. Can possibly. Go wrong? Now we know.
In the end, it's a quality issue. Quality of software, of training data.
"We're trying to understand". Yeah sis, because you didn't pay attention to the supply chain.
It's like you picked a random vendor selling any liquid for your next cocktail and then without checking the label made yourself a nice "vodka" lime. Turns out you bought methanol. Have fun 6ft under.
https://github.com/sunblaze-ucb/exploitgym/commit/e4123d043774623b2274e6bbe0155a423d631f0a
What. Can possibly. Go wrong? Now we know.
In the end, it's a quality issue. Quality of software, of training data.
"We're trying to understand". Yeah sis, because you didn't pay attention to the supply chain.
It's like you picked a random vendor selling any liquid for your next cocktail and then without checking the label made yourself a nice "vodka" lime. Turns out you bought methanol. Have fun 6ft under.