Patent Filed — U.S. Provisional Application

Signal-Correlated AI Firewall That Thinks Before It Blocks

How Delphi Security's multi-layer detection architecture correlates signals across heuristic rules, ML classifiers, and LLM arbitration to stop threats that slip past every other AI firewall.

Delphi Security Research Team · March 2026

Every AI firewall on the market today has the same fundamental problem: their detection layers don't talk to each other.

A heuristic engine flags something suspicious. An ML classifier says it's safe. The system trusts the classifier and lets the request through. The attacker knew this would happen — they crafted their prompt to exploit exactly this blind spot.

We built something different. And we patented it.

The Problem: Independent Layers Are Broken

Modern LLM security systems typically stack detection layers vertically: a pattern matcher, a machine learning classifier, maybe an LLM-as-a-Judge. Each layer makes its own decision independently. If any layer says "safe," the request passes.

This architecture has a fatal flaw. Attackers don't need to beat every layer — they just need to beat one, and that one overrides the others. We call this the independent-layer bypass problem, and it's the single biggest vulnerability in AI security today.

How Attackers Exploit Independent-Layer Systems

Crafted Prompt
  → Heuristic: ⚠ Suspicious
  → ML Classifier: ✓ Safe (overrides)
  → THREAT PASSES

Classifier wins. Attack succeeds.

Our Invention: Signal Correlation Changes Everything

Our patented architecture introduces a fundamentally different concept: cross-layer signal correlation. Instead of letting each layer vote independently, we built a system where earlier detection signals modulate how later layers behave.

Think of it like a doctor's diagnosis. A single symptom might not be alarming. But when a blood test shows one anomaly and a scan shows another, a good doctor doesn't treat them independently — they correlate the signals and escalate. Our system does the same thing, in real-time, at the proxy boundary of every LLM API call.

Heuristic Signal → Modulates → ML Routing → Escalates → LLM Arbitration

Multi-Layer Detection Architecture

The system comprises multiple detection layers that operate as a correlated pipeline rather than independent filters:

1. Heuristic Rule Engine — Deterministic pattern analysis generates suspicion signals that propagate forward to influence downstream classification decisions.

2. Adaptive ML Classifier Routing — Machine learning classifiers receive signal context from upstream layers, adapting their evaluation behavior based on correlated threat indicators.

3. Output Verification Engine — Post-response analysis detects evidence of successful attacks in model output — catching threats that bypass all input-side detection.

4. LLM Arbitration with Signal Context — A secondary LLM evaluates ambiguous cases with full cross-layer signal history, not just the raw prompt in isolation.

Key Innovation

The critical breakthrough is that heuristic suspicion signals modulate ML classifier routing decisions through a deterministic rule engine. This means a prompt that triggers even subtle heuristic flags will be evaluated differently by the ML layer — the system becomes more sensitive precisely when it should be.

Coverage Beyond Prompt Injection

Most AI firewalls focus narrowly on prompt injection. Our patent covers a comprehensive threat taxonomy that addresses the full attack surface of modern AI deployments — including agentic AI, RAG pipelines, and machine-to-machine protocols:

• Prompt Injection — Direct and indirect injection attacks targeting LLM instruction boundaries

• Jailbreak & Bypass — Attempts to override safety constraints and system-level instructions

• Data Exfiltration — Conversational techniques to extract training data, PII, and system prompts

• Agent Goal Hijacking — Attacks that redirect autonomous AI agents to perform unintended actions

• MCP Tool Abuse — Exploitation of machine-to-machine protocols to escalate privileges across agent networks

• RAG Poisoning — Injection of malicious content into knowledge bases to compromise retrieval-augmented generation

Real-Time Threat Topology

The patent also covers a real-time threat topology graph that correlates threats across heterogeneous AI infrastructure components. This isn't just logging — it's a living map of how threats propagate across your AI systems, from individual LLM calls to multi-agent workflows to RAG pipeline interactions.

When an attack targets one component, the topology graph identifies correlated signals across the entire infrastructure, enabling preemptive defense of connected systems before the attack spreads.

Why This Matters

The AI security market is projected to reach $800 million by 2032. Every enterprise deploying LLMs, AI agents, or agentic workflows needs runtime protection that goes beyond pattern matching.

With this patent, Delphi Security establishes foundational intellectual property in signal-correlated adaptive routing for AI threat detection — the core mechanism that makes our AI firewall fundamentally more effective than systems built on independent detection layers.

This isn't incremental improvement. It's a different architecture entirely.

Stats

• Attack surface categories covered: 6+

• Correlated detection layers: 4

• Added latency target: <50ms

Patent Notice: The technology described in this article is the subject of U.S. Provisional Patent Application PROV-002, filed by Delphi Security Inc. The specific implementation details, algorithms, and system architectures are protected intellectual property. This article provides a high-level overview for informational purposes only.