Two numbers, published months apart, describe the same job.
One says a frontier model triages alerts with 98% accuracy . The other says the best of five off-the-shelf reasoning pipelines in a realistic replay misses 40.4% of attacks . Both are honest measurements. They are also, taken at face value, irreconcilable — and the gap between them is the most useful thing I found in this review.
I screened 378 unique records, extracted 32 quantitative comparisons, and graded every one by how fairly it was run. This is an attempt to answer a narrower question: for which security-operations tasks does the measured evidence support using an LLM, and under what conditions does it not?
The short answer: LLMs are genuinely useful as interfaces to structured reasoning and as extractors of unstructured context .…