Detect, respond and recover — entirely on the device.

A multi-pronged detection core

Curated SigmaHQ behavioural rules over ETW and Sysmon telemetry, YARA rules on files and in process memory, ClamAV signatures, and two ONNX scorers — a supervised behaviour model and an unsupervised novelty model. A threat missed by one layer can be caught by another, and agreement across layers raises an incident’s severity. In the final product you choose which engines examine the files during a scan, the dual-LLM reasoning pipeline among them, and the order in which they are used.

Two Llama models, on the CPU

A stock Llama 3.2 1B gives every behavioural incident a first read within seconds. A deterministic router ranks the work. A Llama 3.2 3B fine-tuned for AegisAI writes the authoritative explanation, mapped to MITRE ATT&CK. Analyst-grade narratives, produced without a byte leaving the machine.

Response with a safety floor

Malicious files are quarantined into a protected, restorable store. Malicious processes are suspended, not killed, so their memory survives for analysis. A safety floor refuses to touch critical system processes or validly signed binaries. Every action is audited, and every one can be undone.

Incident controls, checked before they are offered

Each incident offers only the actions that can work on this machine at that moment, decided by the engine and checked again every few seconds. A button for a process that has exited, a signed file, or a protected system folder is never shown. One that applies but not yet is greyed out, with the reason in its tip.

A running process can be paused, resumed, asked to close, force-killed, or force-killed with its whole process tree — children first, and the whole tree refused if any member is one Windows needs. A file can be quarantined and restored to its original path, or deleted together with the files the program wrote, and every file involved is listed with its full path. A file can also be declared safe from its incident, even one no engine can read: what was done is reversed — the quarantined copy restored, the paused program resumed — and its incidents are removed from the journal, the action audit and the logs as if they never happened. A signed program that a behavioural rule matched cannot be declared safe, so a trusted system tool can never be cleared by accident.

Sharper and faster detection

157 curated rules, including AegisAI’s own rule for an unsigned program started from a user’s AppData Roaming folder, a common staging place for malware. The 157 rules in this prototype become many more in the final release of the product, and they keep getting updated regularly. Process-tree corroboration is bounded to the incident’s own time window. A vendor-signed program in its own Program Files folder is no longer mistaken for a renamed system tool. A cache of signature verdicts makes a repeat quick scan about three times faster.

A system scan names the engine and the rule behind every finding and every file it could not examine. A file rescanned successfully or declared safe leaves those lists. A locked system file declared safe by its path is passed by while it stays unreadable, then examined like any other once it can be read.

Honest status, built-in attribution

The product reports reduced capability rather than painting a green tick over a degraded system. Every detection rule’s author is credited wherever its matches appear, as the Detection Rule License requires.