ILION: Deterministic Pre-Execution Safety Gates for Agentic AI Systems
New system operates 2,000x faster than commercial baselines with 143 microsecond latency.
Researcher Florin Adrian Chitan has introduced ILION (Intelligent Logic Identity Operations Network), a novel safety architecture designed specifically for autonomous AI agents. Unlike traditional text-moderation systems that scan for harmful language, ILION evaluates whether an agent's proposed real-world action—such as modifying files, calling APIs, or executing financial transactions—falls within its authorized operational scope. The system employs a deterministic five-component cascade (Transient Identity Imprint, Semantic Vector Reference Frame, Identity Drift Control, Identity Resonance Score, and Consensus Veto Layer) to produce a clear BLOCK or ALLOW verdict. Crucially, it requires zero labeled training data, operates with sub-millisecond latency, and provides fully interpretable decisions.
In rigorous testing on the purpose-built ILION-Bench v2—which contains 380 scenarios across eight attack categories—the system achieved an F1 score of 0.8515, a precision of 91.0%, and a false positive rate of 7.9%. Its mean latency was just 143 microseconds. Comparative analysis revealed that existing commercial safety tools are fundamentally mismatched for this task: Lakera Guard scored F1=0.8087, the OpenAI Moderation API scored a dismal F1=0.1188, and Llama Guard 3 scored F1=0.0105. ILION outperformed the best commercial baseline by 4.3 F1 points while operating approximately 2,000 times faster and with a four-times lower false positive rate. This demonstrates a critical gap in current AI safety infrastructure as agentic systems become more prevalent.
- ILION achieves 91% precision and 0.8515 F1 score on a benchmark of 380 adversarial agent-action scenarios.
- The system operates with 143 microsecond latency, making it 2,000x faster than commercial text-safety baselines like Lakera Guard.
- It uses a deterministic, five-component cascade architecture requiring zero training data, addressing a fundamental mismatch in current safety tools.
Why It Matters
As AI agents gain autonomy to perform real-world actions, this provides a critical, high-speed safety layer that existing content filters cannot offer.