TWO DGX SPARKS, 33 LOCAL MODELS, ONE QUESTION: CAN LOCAL AI ACTUALLY RED TEAM?

Summary from the TLDR InfoSec newsletter

Roei Sherman evaluated 33 local models on a two-node DGX Spark cluster for red teaming capabilities. The evaluation consisted of a security capability that tested whether the models can reason by testing them with a three-phase red team scenario, a red team benchmark to test the willingness and accuracy of model responses, and a coding benchmark consisting of Aider Polyglot plus two custom suites. Sherman found that MoE models consistently performed above their model class across all three benchmarks, with Qwen3.6-35B-A3B (FP8) winning the security capability benchmark, Qwen3-Coder-Next (FP8) winning the coding benchmark, and more mixed results on the red team benchmark.