OpenAI may start blocking using their models to make RL envs on the 20th.
Seems like both a good and bad thing for the data industry, either skill floor is going up or its a race to the bottom with jailbreaks
every CyberGym-E2E task is this:
> here is a c/cpp codebase + location of a memory bug
> find memory bug pls
> fix memory bug pls
this is a useful capability but its not the general defensive security eval people attribute it as
every* ExploitGym task is this:
> here is a memory bug + a bunch of debug info
> turn it into code execution pls
this is a useful capability but its not the general offensive security eval people attribute it as
@stephenx_ I can nearly always rewrite LLM text to be easier to understand, more concise while conveying the same (or better) meaning.
Why do I need a precise definition to call the original slop?
@scaling01 exploitgym only rewards for full ace whereas exploitbench gives partial rewards for escalation levels. Also exploitgym is mostly binary exploitation and exploitbench is all v8