608 multiple-choice questions on real malware detonation reports from CrowdStrike's CyberSOCEval (part of CyberSecEval 4), imported from inspect_evals. Each Case pastes one sandbox report, about 131,000 characters on average and up to 260,000, then asks a question where several options may be correct; the model lists every letter it thinks is right. Grading is the eval's own scorer, so no judge tokens are spent: a Case scores 1 only when the letters match the key exactly, and its Jaccard overlap with the key is published beside it as partial credit. Benchmark score = exact-match accuracy over the cases run; the eval's per-topic, difficulty and attack breakdowns are not reproduced. One question with no answer key is left out. inspect_evals runs the model at temperature 0; this benchmark imposes no generation settings. Use a model with a long context window. Answer format: end the reply with a line like 'ANSWER: B, D' that names the letter of every right option and no others; 'ANSWER: B and D' also works, and so does each letter wrapped, as in 'ANSWER: (B), (D)'. Naming only some, as 'ANSWER: B' or 'ANSWER: (B)' does when B and D are right, scores 0.