Industry first
No endpoint-protection product on the market today lets its customer take part in training the language models. AegisAI does. The owner — the person at the machine, or in AegisAI Local — SME Edition the company’s own security team — adds knowledge on this machine. What they add builds on the fine-tuning AegisAI ships and keeps current. It never replaces it.
Out of the box both models know Windows threats in general and nothing about your organisation: which administration tools you sanction, which scripts IT runs at night, where your own software lives, which practices your policy forbids, and whom to call. So an explanation can only say what happened, not whether it fits the way you work.
Fine-tuning teaches them that knowledge. An explanation can then say “this is the nightly backup job your policy allows in section 4.2, started from its approved folder” instead of “an unknown script ran with administrator rights”, and the first read of an incident reflects what is routine here. You spend less time on behaviour you have already sanctioned, and every explanation is written against your own rules.
LoRA (low-rank adaptation) does this without retraining a model. It learns a small adapter file, tens of megabytes, that sits on top of the unchanged base model. An adapter can be tested beside the one in use before it is switched on, and switched off again by going back to the previous file.
Tier 1 — triage
In this prototype that model is Llama 3.2 1B. In the finished product it is an always-loaded Llama 3.2 3B, about 1.7 GB and about 80 ms, handling about 95% of events. An adapter teaches it your vocabulary and what is routine on your machines, so its first read, and the order in which the router hands incidents to the deep model, reflect your environment. Fine-tuning buys speed and calibrated confidence.
Tier 2 — deep explanation
In this prototype that model is Llama 3.2 3B with the AegisAI adapter. In the finished product it is Llama 3.1 8B on demand, about 4.5 GB and about 1.2 s cold start, for the uncertain and novel few. Your knowledge is added to the AegisAI training, not put in its place: the new adapter learns from the AegisAI examples and yours together, so the security knowledge the model already has is kept. Fine-tuning buys grounding and storyline fidelity.
From a policy document to a tested, reversible adapter, without anything leaving the machine.
1
Upload a personal or company security policy, or pick incidents you have already reviewed and write the explanation each one should have had.
2
AegisAI parses each item into lines of a JSONL training file, one example per line, in the exact form each model reads when it works: the same fixed instructions, an event window drawn by the product’s own renderer, and the answer the model should give. Training and use see the same bytes, which is what makes an adapter work.
3
Every extracted item is listed with the place it came from. You edit its wording, keep it or drop it, and choose which model receives it: the triage model, the deep model, or both. Only what you keep is imported.
4
A LoRA adapter is trained for each chosen model on its full-precision base, on this PC or on a training machine you name. Only the adapter’s weights are learned. The base model does not change. On the 4-core development machine the triage model learned 250 examples in about five and a half hours. The deep model is best trained on a machine with a GPU.
5
The new adapter is tested on examples held back from training and on the product’s golden test window. It must keep the exact answer format and do at least as well as the adapter in use. One that does worse is refused, and the page shows where it fell short.
6
An adapter that passes is converted to the format the engine loads (GGUF) and imported into its model in the pipeline, taking effect at the engine’s next start. The previous adapter is kept, so going back is one choice on this page.
A personal or SME security policy can be uploaded as PDF, Word, Markdown or plain text. It is parsed on this machine and nothing is sent anywhere. AegisAI splits it into statements — an approved tool, a sanctioned script, a forbidden practice, an exception, an escalation contact — and turns each one into training examples that show the policy applied to events, not just its text, so a model learns to recognise the situation the policy describes.
A stock language model knows English, not endpoint telemetry. Sysmon event IDs, parent-child process chains, obfuscated command lines and the exact boundaries between ATT&CK techniques are a specialist grammar, and a general model reads them the way a well-read layperson reads an X-ray. Fine-tuning on real, labelled telemetry turns the finished product’s Llama 3.2 3B triage model and Llama 3.1 8B deep model into specialists.
| Figure 5 | Stock general model | Fine-tuned AegisAI specialist |
|---|---|---|
| Reads telemetry | Like a well-read layperson | Like a trained analyst |
| Grounding | Fills gaps with plausible stories | Cites what the rule asserts; says “not observed” where evidence stops |
| Confidence | A number with nothing behind it | Calibrated on labelled outcomes — safe to act on |
| Every prompt carries | Instructions, format, grounding rules, and the evidence | The evidence alone, in a compact encoding |
| Output | Varies from run to run | The same structured verdict for the same evidence |
| Size it needs | Larger, to compensate | Small enough for an ordinary PC |
Figure 5 — what fine-tuning changes. The same base model, before and after it is trained on real, labelled endpoint telemetry.
01
An untuned model fills gaps with plausible stories: a sibling technique instead of the right one, a command-and-control beacon inferred from a single encoded command. Trained on windows labelled against their rules and techniques, the model cites the technique the rule actually asserts and says “not observed” where the evidence stops. Whole techniques are held out of training, so this is measured on attacks the model has never seen.
02
In the finished product the triage tier sends its confident majority, about 95% of cases, straight to the response engine. That is safe only if “95% confident” means right 95% of the time. A stock model’s stated confidence is calibrated against nothing. Fine-tuning on labelled outcomes is what makes the number mean something — it is what moves AegisAI from explaining to acting.
03
On a CPU, reading the prompt dominates the time to an explanation — about 94% of it, measured on this prototype’s deep tier. A general model has to be told, on every call, the instructions, the output format and the grounding rules. A fine-tuned model carries them in its weights, so the prompt shrinks to the evidence alone, in a compact encoding the model was trained to read. Every token removed is time saved, almost linearly, and a triage tier trained to answer in a fixed structured form needs only a few dozen output tokens.
04
A tuned model returns the same structured verdict — technique, severity, confidence, storyline — for the same evidence. The two tiers can then be measured against each other on every incident, and AegisAI already records where they disagree. Quality becomes a number the product watches, not an impression.
05
Published results repeatedly show small models tuned to a narrow, well-defined task matching much larger general models on that task. That is what lets AegisAI’s reasoning run on an ordinary PC, with no GPU and no data centre. A bigger stock model would cost memory and latency, and would still be uncalibrated on telemetry.
Triage — Llama 3.2 3B
Fine-tuning buys speed and calibrated confidence. The gatekeeper.
Deep — Llama 3.1 8B
Fine-tuning buys grounding and storyline fidelity. The specialist.
Figure 6 — two specialists, one gated learning loop. The loop repeats every cycle and keeps earlier data, so nothing already learned is forgotten. An adapter is promoted only after held-out testing and shadow evaluation.
Beside the models
The same ownership extends to every detection engine. Personal Sigma behavioural rules, YARA rules for files and process memory, ClamAV-format signatures and other detection content can be written in place or imported, tested on the machine against its own telemetry before they are switched on, and kept apart from AegisAI’s updates. Together with the trained models, AegisAI is the first endpoint product whose detection logic and whose AI are both shaped by the customer.
