Measuring LLMs' ability to develop exploits / hacker news