Decision models could reshape AI SOC alert triage, but context remains the bottleneck

A new class of AI decision model has emerged in the past month, with TypeSafe AI releasing Jev on 15 September, OpenAI opening a public beta of its Decisions API on 6 October, and Cribl publishing the open cribl-decision LoRA adapter. These models return a typed choice, score or probability rather than generated text, which makes them well suited to automated alert triage.
The economics are compelling: Jev is priced at $42 per billion input tokens with 70 to 500 millisecond responses, putting the annual model bill for an enterprise handling 450,000 alerts at roughly $100. That removes the case for only inspecting high-severity alerts, and every verdict becomes an auditable record.
The article warns that calibration is only as good as the input. In data covering 25 million alerts triaged in 2025, 51% of machines with an active infection had already been reported as mitigated by EDR, and nearly 1% of real incidents traced back to the lowest severity alerts. Cribl’s own model card notes the model can be overconfident and is sensitive to prompt wording.
The author argues the missing layer is Security Context Engineering: collecting process trees, command lines, memory forensics and identity context before a verdict is made. Organisational context cannot be bought and decays quickly, so it must be maintained as a loop that feeds outcomes back into the model.