here’s what i know about how an openai model hacked a vulnerable test system. it all begins with this one paper named ExploitGym.
one of the tasks was to exploit the memory system in the v8 engine. the vulnerability was a memory bug that allowed the model to read a memory address, or pointer, from a v8 object.
the function they wanted was puts, not PUT. from the address of puts, they could find the base address of libc, and from libc, they found system().
with system(), they could run shell commands, such as sending a request to a url using curl.
Top comments (0)