Security context system outperforms coding agent in cloud investigation study

In a preprint comparing cloud security investigations, researchers report that the Sola Security Brain covered more findings than Claude Code across 28 tasks in the same live AWS environment. Its coverage score was 0.693, versus 0.387 for the coding agent, with 17.7 times lower reasoning cost per task. The system resolves relationships among cloud resources before answering queries, while the coding agent investigates through a read-only command-line interface. The researchers describe one case in which the agent reported that no bucket policies existed after sampling 40 of roughly 5,000 buckets, although 65 buckets had a wildcard-principal read grant. The results come from the study's task set and grading method.