A new category of AI model, the "decision model," has gone from nowhere to everywhere in a month. Rather than generating text, these models take a state and typed questions, then return a choice, score, or probability. TypeSafe AI released Jev on September 15, OpenAI launched its Decisions API in beta on October 6, and Cribl published an open LoRA adapter. For security operations, the appeal is obvious: a calibrated probability can be automated against thresholds—close benign alerts above 0.97, escalate below 0.6, and let analysts handle the middle.
The economics help too. Jev is priced at $42 per billion input tokens with responses in 70–500 milliseconds. An enterprise handling 450,000 alerts a year, feeding a few thousand tokens each, would spend roughly $100 annually. That makes it feasible to triage every alert, not just the high-severity ones. And because outputs are schema-fixed with probabilities, every verdict can be logged, replayed, and audited.
Still, a well-formed answer can be wrong. Calibration is a distribution property; it does not guarantee correctness on any single input. The article's example: an alert reading "malware detected, threat quarantined, low severity" may usually be fine, but in the author's data over half of infected machines had already been reported as mitigated. Cribl's model card is similarly honest, warning that the model can be overconfident and sensitive to wording. The input, not the model, often carries the mistake.
The fix, the article argues, is what it calls Security Context Engineering: gathering process trees, memory scans, login history, and other forensic evidence before making a call. The hardest context is what "normal" looks like for a specific organization—identity alerts show this plainly, with impossible travel rarely a real compromise and VPN usage causing many false positives. Decision models are a useful tool, but they only see what they are shown. A calibrated guess is still just a guess.