Expert Calls For Maximal Performance Testing Of Frontier AI On CBRN Risks
Reactions from ranked influencers
2 posts
I know I am a broken record at this point but to assess frontier capabilities, especially for CBRN, we need to try everything to maximize performance. This means finding the harness that maximizes performance for any model. Bad actors won’t write ReAct loops.
And get rid of token limits for frontier agent evals, AISI (and MirrorCode and EdgeBench and and and) have shown that we need to increase the limits by a lot to see the ceiling. Assessing (cyber) risk means to push as hard as possible, which some benchmarks cant show
OpenAI’s GPT-5.5 is the second model to complete one of our multi-step cyber-attack simulations end-to-end 🧵