Overview
ChimeraGuard and NetMon were built around a single architectural principle: no consequential security decision should rest on a single evaluator. The system that emerged runs two completely independent layers against every network event — a deterministic rule engine and a local AI reasoning layer — and treats their disagreement as the most informative output of all.
Most detection systems either commit fully to rules (fast but rigid) or commit fully to AI (context-aware but uncertain). The hybrid approach accepts that both are right about different things. Rules are reliable on the threats they know. AI is useful on the context rules cannot see. Neither is sufficient alone, so neither is trusted alone.
At a Glance
- Built for: real-time network security monitoring on live systems
- Core systems: NetMon rule engine + ChimeraGuard AI layer + baseline suppression
- Key outcome: security alerts that combine rule precision with contextual narrative
- Resilience: AI failure produces a labelled fallback, never a silent gap
Features
Ten-Rule Deterministic Engine
Every connection event passes through ten rules before any AI is involved. The rules cover the threat categories that produce the clearest signal: connections to known command-and-control ports, privileged system processes making external connections they should never make, scripting interpreters connecting outbound, processes running from temporary or user-writable paths, executables with non-standard extensions, destinations flagged by VirusTotal, IPs with high AbuseIPDB abuse scores, connections to Tor exit nodes and VPN infrastructure, connections with no reverse DNS and no recognisable organisation, and unrecognised processes connecting externally with no enrichment data. Each rule carries a severity — critical, high, medium, or low — and multiple rules can fire on the same event simultaneously.
AI Reasoning Layer
When a rule fires, the full connection context — process identity, executable path, command line, local and remote addresses, protocol, reverse DNS, organisation, ASN, country, Tor and VPN flags, VirusTotal engine count, and AbuseIPDB score — is passed to a locally running language model for independent analysis. The model is asked to explain what is happening in plain language, assess the threat level, and recommend one clear action. It does not see the rule verdict before forming its own assessment. That independence is the point.
Task-Routed Model Selection
ChimeraGuard does not use a single model for everything. Fifteen task types each have a designated primary model and a fallback model. Fast triage tasks go to a smaller, lower-latency model. Deep analysis and reasoning tasks go to a larger model where output quality matters more than speed. When the primary model is unavailable, the system automatically routes to the fallback rather than waiting or failing.
Three-State Baseline Suppression
NetMon's baseline system operates in three explicit states. In learning mode, every connection seen on the network is recorded as a known-good pair — which process connected to which remote address. In active mode, any event matching a recorded pair is suppressed before the rule engine even runs. This means the rule layer sees only genuinely new or unexpected behaviour rather than re-evaluating known-good infrastructure on every cycle. The transition between modes is manual and intentional.
Labelled Fallback on AI Failure
If the local model is unreachable, times out, or returns an empty response, the system produces a rule-based fallback response automatically. The fallback is clearly labelled so that analysts reviewing alerts know whether they are reading AI-generated context or automated rule output. A partially informed alert surface is better than a confident-looking gap.
What It Resolved
Before the hybrid architecture, the choice was binary: rules that fired reliably but produced no context, or AI that produced context but could not be trusted without a rule anchor. Pure rule systems created alert queues where analysts had to manually look up every destination, research every process, and determine context themselves. Pure AI systems created uncertainty about whether the model's assessment was grounded in real threat intelligence or a confident-sounding hallucination.
The hybrid resolved both problems simultaneously. Rules provide the trigger and the evidence anchor. AI provides the narrative that makes the alert immediately actionable. Disagreement between them — a rule firing on something the AI assesses as benign, or AI escalating something a medium rule caught — surfaces cases that need human attention precisely because automation is uncertain.
The baseline suppression resolved the second major problem: alert volume from known-good infrastructure. Without suppression, routine traffic from patch management, monitoring agents, backup jobs, and cloud sync clients generates constant rule firings that analysts learn to ignore. The baseline layer eliminates that noise systematically without requiring manual exception lists that decay as infrastructure changes.
Usage
NetMon runs continuously on the host system, collecting live connection data and evaluating every external connection against the rule engine. Alerts appear on a dashboard with the process identity, connection details, destination intelligence, rule verdict, AI assessment, and recommended action in one view. Analysts can review, confirm, or dismiss alerts. The baseline system is engaged when the network environment is stable enough to learn.
ChimeraGuard operates as the AI layer that NetMon invokes for alert analysis. It is also accessible as a standalone reasoning interface through Jarvis Nexus, where its model routes handle tasks ranging from triage and alert analysis through deep investigation, code inspection, dossier generation, and correlation across multiple alerts.
Benefits
- Reduced analyst workload: every alert includes a plain-language explanation and a recommended action, eliminating the manual research step that slows triage.
- Resilience to AI failure: the system continues functioning at rule-layer precision when the local model is unavailable, with no silent degradation.
- Transparent uncertainty: disagreement between layers is surfaced rather than resolved automatically, directing human attention where automation is genuinely uncertain.
- Lower alert volume: baseline suppression eliminates known-good traffic from the rule engine, making remaining alerts meaningful by default.
- No cloud dependency: the entire AI layer runs locally on Ollama, with no data leaving the host for inference.
- Versioned and auditable: rule version, model name, threat level, and fallback status are stored with every alert record.