Platform
One detection substrate under every surface
Most AI security products bolt a rule engine onto each connector, so a detection has to be written five times and still behaves differently in each. The Detection Fabric is one asynchronous scoring layer that every signal crosses — whatever surface it came from — and one incident store that everything correlates into.
Capabilities
Key Features
One substrate, every surface
Copilot, Gemini, OpenAI, Anthropic, cloud, gateway, endpoint and your own agents all cross the same scoring layer. A detection written once applies everywhere, instead of being reimplemented per connector.
Asynchronous by design
The fabric is a decoupled consumer reading from the log store. It never sits on the ingest write path, so detection cannot slow, block or drop your telemetry — and a detection outage is not a data-loss event.
Routed, not scanned
An inverted index selects candidate rules per signal instead of evaluating the whole catalogue. Per-signal cost stays flat as the rule count grows, which is what makes a large catalogue affordable.
A cascade, not one model
Cheap deterministic checks run first, statistical baselines next, graph reasoning after that, and a model only where a model is actually needed. Most signals are resolved before anything expensive runs.
Continuous, with durable position
Each engine keeps a watermark in the store and saves it only after the write lands. A restart re-reads a small overlap rather than losing a window — crashes are harmless, never silent gaps.
It defers to a real classifier
Where an upstream scan has already judged a signal with a model, the heuristic layer stands down instead of re-flagging what the classifier cleared. Fewer contradictions, fewer false positives.
The cascade
Four tiers, cheapest first
Each tier answers something the one before it structurally cannot. Running them in cost order means the expensive reasoning only sees what actually needs it.
Stream
detection-engine Signal by signal
Canonicalise, tag, accumulate across multiple scopes, route to candidate rules, evaluate, correlate. Validated classifiers do the tagging — Luhn-checked cards, provider-signature secrets, Trojan-Source and zero-width detection — so a finding survives review.
CatchesPrompt injection, secrets in output, PII disclosure, ASCII smuggling
Behavioural
analytics-worker Against its own baseline
Online baselines per identity and per agent, updated incrementally, scored by deviation. UEBA and AEBA over request rate, tool diversity and data access — no threshold anyone has to guess at and then maintain.
CatchesAnomalous agent behaviour, sudden tool-scope expansion, insider patterns
Graph
graph-correlation Structure no event reveals
Recomputes reachability over the Context Graph and looks for structural conditions: toxic combinations, and multi-stage campaigns that only exist as a shape across several entities.
CatchesToxic combinations, agent-to-agent propagation, exfiltration paths
Deepfake
media-analyzer Synthetic image, audio and video
Provenance and manipulation analysis across image, audio and video — C2PA and EXIF first, a calibrated ensemble only where provenance is silent — feeding verdicts into the same incident pipeline.
CatchesDeepfake and synthetic-media risk in AI workflows
Why it is built this way
The constraints that shaped it
Detection must never break ingestion
Inline detection means a slow rule becomes a dropped log. The fabric reads from the store instead, so the worst case for a detection failure is delayed findings — never missing telemetry.
A catalogue only grows
Evaluating every rule against every signal makes each new detection a tax on throughput. Routing through an inverted index means the hundredth rule costs what the first one did.
A model on every log is unaffordable
And unnecessary. Deterministic checks resolve the overwhelming majority. The cascade exists so judgement is spent where judgement is actually required.
Restarts are normal
Watermarks persist after the write, not before, and resume with a deliberate overlap. Reprocessing is idempotent, so a crash costs a little duplicate work rather than a hole in your coverage.
The fabric produces incidents, not alerts. Everything it finds lands in the same store the Agentic AI-SOC works out of, keyed to the same entities as the Context Graph — including verdicts from Deepfake Detection, so a synthetic asset correlates with whoever introduced it and whatever it reached.
Get started with Detection Fabric
See how Saf3AI can help secure your AI agents.