The CognitionAI Engine™
The CognitionAI Engine neutralizes AI-driven attacks before they execute — at machine speed, with no human in the loop. Neuromorphic mathematics, not a language model. It powers every Salience Cyber product, and it is the architectural reason we can act before execution.
A prevention engine, not a detection tool.
Conventional tools ask "has this happened before?" — matching signatures, behavioral baselines, and post-event telemetry. They were built for attacks that moved at human speed and left recognizable artifacts.
The Engine asks "is this about to happen?" One decision system spans the browser plane and the system plane, predicts what comes next, and acts in microseconds, in-process — no analyst, no playbook, no signature lookup.
- ANeuromorphic, by design
Computation modeled on biological neural systems — sparse, event-driven, predictive. Not another layer stacked on conventional ML.
- BPurpose-built for AI-driven threats
Built from first principles for AI-Enhanced, AI-Native, and AI-Generated attacks — classes legacy stacks have no language to describe.
- CAutonomous, no kernel agent
The userspace System Sensor and an AI-aware Browser Sensor, intercepting agentic AI where users meet it. No analyst grading alerts, no rip-and-replace.
- DTwo planes, one decision loop
The browser plane (where teams meet AI) and the system plane (where the attack lands) feed one decision system — not two tools trading alerts.
Grounded in neuromorphic mathematics.
Two things sit under everything the engine does. Neuromorphic computation— sparse, time-encoded and event-driven, rather than dense matrix multiplication on a fixed graph. And neuromorphic semantic reasoning — reading what an interaction means, not what it matches. Together they are why the engine can commit on the leading edge of activity instead of after a window closes.
Intent is a prediction, not a guess at motive. Activity on both planes is encoded as event spikes; the engine builds a sparse representation of the sequence and predicts its continuation — what it is on course to do, scored before the next step runs, with no prior sample required.
The method is Hierarchical Temporal Memory and Sparse Distributed Representations, modelled on how the neocortex represents sequences. A prompt or tool call enters as a sparse encoding, not as tokens for a model to read — so there is no language in the decision path for an attacker to address.
The output is explainable: the engine accounts for a neutralization in plain language, not a score an operator has to take on trust.
Activity on both planes is encoded as event spikes.
Sparse representations score vector, exposure, and severity against the predictive model.
The attack class is read from pattern semantics, not matched to a known sample.
The engine projects the next step. The window of exposure closes before it opens.
Action is the same computation as the prediction. No handoff, no ticket, no human.
| Salience Cyber | Model-judged AI security | Post-collection analytics | The stack you already own | Policy only | |
|---|---|---|---|---|---|
| Where it sits | In-line, at the point of interaction | In the request path, as a proxy or gateway | On a server, after the fact | On the host and in the log, after execution | Nowhere. It is a document |
| What it reads | Intent | A model's reading of the text | Logs | Artifacts and behaviour | Nothing |
| Can the decider itself be talked to? | No language in the decision path to address | It reads attacker-controlled text by design | Not in the path; it does not decide | Not in the path; it decides after execution | The policy cannot. The person can |
Same words. Different architecture. See the full comparison →
What it takes to run.
Catches what has never been seen.
Sparse representations generalise to patterns never seen before — no signature dependency, for AI-generated malware, prompt injection or agentic exploits.
Microseconds, because nothing leaves the process.
The loop closes before the OS schedules the process or the injection lands — pre-execution, not post-mortem.
Updates itself in production.
Plasticity adapts the engine in place — no model to retrain, roll out, or label. Time-to-protect is engine latency, not release-train latency.
Zero analyst grading load.
Decision and action are one computation: no alert queue, no playbook, no escalation. No dependency on GenAI, LLMs, SLMs, or MCP gateways, and zero token cost — a flat, predictable TCO.
Day-one, no kernel agent, no rip-and-replace.
Runs standalone or beside your existing SIEM, EDR and SOC. Both planes, day one.
All three classes of AI-driven attack.
AI-Enhanced, AI-Native and AI-Generated — one decision system, not three products forwarding alerts to a fourth.