Semantic Threat Intelligence

Mail security that thinks in embedding space.

VecNet replaces brittle signature-based filtering with a learned semantic boundary. Every inbound message is projected into a 768-dimensional latent space and scored against a proprietary threat manifold before it ever reaches a mailbox.

Open Webmail Read the whitepaper →
benign_cluster.centroid threat_cluster.centroid decision boundary
// EMBEDDING GATEWAY

Inline vector encoder

Every message body, header, and attachment fingerprint is passed through VecNet's proprietary sentence encoder before the SMTP transaction is allowed to complete. Nothing touches disk unencoded.

// THREAT MANIFOLD

Continuously retrained

The manifold is refit nightly against telemetry from every tenant on the network, so the boundary between benign and malicious semantics shifts as attacker language evolves.

// NEAREST-NEIGHBOR AUDIT

Explainable scoring

Flagged mail ships with its k-nearest neighbors in latent space, so analysts can see exactly which historical vectors informed the decision — full provenance, zero black box.

768d
latent representation
14M
indexed embeddings
99.2%
manifold coverage
<40ms
inline scoring latency
Restricted Endpoint

Tenant webmail is gated behind credentialed access to the mail server. Provisioned accounts authenticate directly against the mailstore — there is no separate identity layer in front of it.

servicevecnet-mailstore
protocolIMAP / SMTP (AUTH required)
statusONLINE
notecredentials are not self-service