We burned 11.7B tokens to find the best cyber AI model

LoganDark

Of course, the model trained most recently does the best job. Do we know what DeepSeek V4 Pro 0813's knowledge cutoff is? For all we know, it's simply working in parallel with how the vulnerabilities were first discovered, even if the solutions themselves weren't trained in. Surprised to see Grok 4.6 performing so well. Shame I can't use it privately -- I would never risk getting my Twitter account banned for that

exceptione

I don't see Fable, but I guess Fable would refuse to work anyway on cyber problems(?) But why would Opus 5 work then?

k4roshi

How was this so urgent that it required sleepless nights for the team?!

joker99

Since the “magic“ is in the harness: can anyone recommend a good OSS cyber harness? I’ve been experimenting a bunch at $work and for large heterogeneous code bases, just using codex or claude seems to work better than experimenting with tools like code graph/graphql to save on tokens