There is a language model somewhere in most security products shipped in 2026. Often it is genuinely useful — translating a natural-language question into a query, summarising an incident, drafting a report. We use language models ourselves for exactly those kinds of tasks.
But there is one place in AegisCore where you will not find one: the Decision Engine. The layer that scores risk and recommends action is fully deterministic, by design. This post explains why.
A decision is not a summary
A summary can be approximately right and still be useful. If a model paraphrases an incident and gets a detail slightly wrong, an operator notices and moves on.
A decision is different. When the platform concludes "this host is compromised, isolate it" or "this finding is low risk, deprioritise it," that conclusion drives an action. It needs to be reproducible — the same inputs must always produce the same output — and it needs to be explainable down to the specific rule and value that produced it.
A language model offers neither guarantee. It is probabilistic by construction. Run the same prompt twice and you may get two answers. Ask it why, and it will generate a plausible-sounding rationale that may or may not reflect what actually happened inside the model.
The hallucinated CVE problem
Here is the concrete failure mode. During evaluation of an LLM-assisted security tool, we watched it reference a CVE identifier that does not exist. It cited a source — a well-known vulnerability database — that had never published anything under that identifier. The model had invented a vulnerability, attributed it to a real authority, and presented it with complete confidence.
In a research context that is an annoyance. In a SOC 2 audit, where an auditor is reviewing why your platform blocked an IP address or escalated an incident, a confidently invented metric is a finding against you. You cannot defend a decision your own tool cannot reconstruct.
What deterministic actually buys you
AegisCore's Decision Engine scores risk through an explicit framework: defined inputs, defined weights, defined thresholds. The same evidence always yields the same score. When the engine recommends an action, the D3 explainability dashboard shows the decision tree that produced it — the exact path, the exact values, no reconstruction required.
This has three consequences that matter:
- Audit survives contact with reality. Every decision maps to 50 SOC 2 controls with signed evidence. An auditor inspects the artifact, not our word.
- Drift is controllable. Because the logic is explicit, a change in behaviour is a change in code — visible, reviewable, version-controlled. There is no opaque weight update that quietly shifts how the platform reasons.
- Operators can trust it. The console explains itself. An analyst can disagree with a decision, trace exactly why the engine made it, and override it with full context.
Where language models are welcome
This is not an anti-LLM argument. It is an argument about where probabilistic systems belong.
Outside the decision path, language models do real work in our world — generating content drafts, researching context, translating queries. Those tasks share a property: a human reviews the output before it has consequences, and an error is visible. That is the right place for a probabilistic tool.
The decision path is not that place. When the output drives an automated containment action or lands in front of an auditor, "approximately right" is not a specification.
The honest version of "AI-powered"
A lot of security marketing in 2026 reduces to "we added an LLM." Sometimes that genuinely helps the operator. Often it just moves the hallucination risk closer to the decision.
AegisCore's position is narrower and, we think, more honest: the decision layer is deterministic and explainable; language models assist around the edges where a human stays in the loop. If you are buying a platform whose conclusions an auditor will scrutinise, ask exactly where the language model sits. The answer should not be "in the decision."
You can explore the Decision Engine yourself in the free Community edition — see the download page.