XBOW’s sponsored writeup says Mythos Preview cut false negatives by 42 percent on its web exploit benchmark, and by 55 percent when both models had source code. The useful caveat is that the same article says live-site access still mattered more than source access for proving bugs, while command-safety judgment lagged some baselines at 77.8 percent.