Flipping the eval on its head
A new approach to cybersecurity evaluations proposes using language models as constant red-team oracles to empirically compare the attack surfaces of different software implementations, such as OpenSSL forks or formally …