Tweet by Siméon:

If an LLM scores 40% on Cybench, what does that tell you about its ability to help in a real-world actor at cyber offence?

1. We just released a first paper proposing a method to get an answer to that question.

2. Cybench is particularly suited to real-world because it has a… https://t.co/FWBXX7cOgD pic.twitter.com/s4Fk8RW0e8