Defender Impact
Decision models represent a structural shift in how AI can be applied to alert triage: instead of generating explanatory text that analysts must interpret, they return typed, probabilistic answers that automation can act on directly. For the first time, the economics and latency of AI-assisted triage make it viable to apply a verdict to every alert — not just the ones that survive severity filtering.
Capability Overview
Within a single month, a new model category materialised across multiple vendors. TypeSafe AI released Jev on September 15, framing it as a ‘System One’ model — fast, intuitive, and non-generative. OpenAI announced its Decisions API at DevDay and opened a public beta on October 6. Cribl published cribl-decision, an open LoRA adapter designed for local deployment. Others, including Laya, are appearing in open-source tooling.
The core mechanic is a deliberate departure from language model design. You supply a state object and a set of typed questions. The model returns a choice, a score, or a yes/no with a calibrated probability attached — not a paragraph. This schema-bound output means the answer always conforms to the structure you defined. Vendors correctly note this eliminates hallucination in the traditional sense: the output will always be well-formed. What it does not eliminate is the possibility of a well-formed wrong answer.
Pricing reinforces the appeal for security operations. Jev is priced at approximately $42 per billion input tokens with 70–500ms response times. At 450,000 alerts per year with a few thousand tokens per alert, the annual model cost is roughly $100. That number effectively dissolves the economic justification for ignoring lower-severity alert volumes.
Defensive Advances
Decision models open several concrete advances for security operations teams:
- Universal triage coverage: Every alert — not just high-severity ones — can receive a probabilistic verdict cheaply enough to act on. The long-standing practice of discarding lower-severity alerts for capacity reasons becomes an operational choice rather than an economic necessity.
- Threshold-driven automation: A calibrated probability score maps directly onto runbook logic. Auto-close above 0.97; escalate below 0.60; route to analyst in between. Critically, TypeSafe’s guidance to apply stricter thresholds to irreversible actions (disabling an account vs. adding a comment) gives teams a practical framework for calibrating automation risk.
- Auditable decision records: Because the answer space is fixed and the output carries its own probability, every triage decision can be logged, replayed, and compared against ground truth. This is materially easier to audit than free-text reasoning chains.
- On-premises deployment path: Cribl’s open LoRA adapter enables air-gapped or regulated environments to run decision model inference locally, extending the benefit to sectors where cloud API calls are constrained.
Residual Gaps
The capability’s value is directly proportional to the richness of the context provided to the model. Calibration is a promise about a distribution — when a model says 90%, it means that across inputs that look like this one, it is right 90% of the time. The operative phrase is ’look like this one.’
Data from Grid the Grey’s AI SOC Report 2026 illustrates the stakes: across 25 million triaged alerts, 51% of infected endpoints had already been reported as mitigated by the EDR at the time of alert. A decision model fed only an alert title and severity field would see ‘malware detected, threat quarantined’ and have no mechanism to surface the false-mitigation signal hiding in live host memory state.
Key maturity questions organisations should resolve before scaling deployment:
- Context completeness: Are host state, live EDR telemetry, network context, and historical alert patterns included in the input, or just alert metadata?
- Calibration validation: Has the model been validated against your environment’s specific alert distribution, not just the vendor’s benchmark corpus?
- Threshold tuning: Have thresholds been tested against historical true/false positive rates before live automation is enabled?
- Integration maturity: Does your SIEM or SOAR support typed probabilistic inputs natively, or does middleware need to be built?
Framework Mapping
Decision models most directly address LLM09 (Overreliance) concerns by making AI confidence explicit and threshold-governed rather than implicit in generated text. They partially address AML.T0015 (Evade AI Model) risk by narrowing the answer space, reducing the surface for adversarial input manipulation — though evasion via crafted alert metadata remains a consideration for mature deployments. LLM02 (Insecure Output Handling) risk is substantially reduced by schema-bound outputs eliminating free-text injection paths into downstream tooling.
Deployment Considerations
Organisations should sequence adoption in three phases. First, run the model in shadow mode against historical alerts to validate calibration against your own data distribution. Second, enable auto-close only at high-confidence thresholds for reversible actions, with full logging active from day one. Third, progressively enrich the input context — adding host telemetry, EDR state, and asset criticality — before expanding automation to lower-confidence bands or higher-impact actions.
Teams operating in regulated environments should evaluate Cribl’s cribl-decision LoRA adapter as an alternative to cloud API deployment.
Defender Checklist
- Identify your current alert volume and estimate annual decision model cost against your triage backlog
- Define threshold tiers mapped to action reversibility before any live automation is enabled
- Audit the context fields in your planned model input — confirm host state and EDR telemetry are included
- Run shadow-mode validation against 90 days of historical alerts before enabling auto-close
- Establish a weekly calibration review comparing model verdicts against analyst-confirmed outcomes
- Evaluate on-premises LoRA adapter options for environments with cloud API restrictions
- Log every model decision with its probability score for future audit and model drift detection