Continuous-time defense · BEC · VEC · account takeover

Every email passed. The thread didn’t.

Attackers hijack real vendor mailboxes and send clean, payload-free emails that pass SPF and DKIM. We model the whole conversation as a temporal graph and stop the takeover while it’s still talk, in your tenant, in under 50 ms, retaining nothing.

Join the pilot waitlistRead the researchWatch an attack unfold
<50 msper-message inference, in-tenant
Zerodata retention
1–3Bparameter quantized intent SLM

The problem

Authentication passed. Trust was already stolen.

Gateways and frontier LLMs judge each email on its own. A multi-turn compromise never looks bad in any single message, which is exactly why it works.

Legacy · transactional

One message in, one verdict out

  • Passes SPF, DKIM and DMARC because the mailbox is genuinely compromised.
  • No payload, clean grammar, calm tone: nothing for a classifier to flag.
  • Context window is one email, so trust built over turns is invisible.

T1 PASS → T2 PASS → T3 PASS

Continuous-time · relational

The thread is the unit of analysis

  • Models the tenant as a dynamic graph: who talks to whom, when, with what authority.
  • Bochner harmonic time encodings expose reply-rhythm anomalies.
  • Flags new participants, bypassed approvers and authority inversions.

T1 OBSERVE → T2 WATCH → T3 QUARANTINE

Live demo

Watch a vendor email compromise unfold, one turn at a time.

Step through four emails. Left: what a transactional filter concludes. Right: what the graph engine sees.

Gateway / frontier LLM

Inspects one message at a time

Clean (Pass)

    TGAT + Intent SLM · in-tenant

    Reads the whole thread as a graph over time

    INFERENCE LATENCY

    0 ms / 50 ms budget

      Illustrative simulation with synthetic data. Not a live detection trace.

      Under the hood

      Graph attention for the relationship. A small language model for the ask.

      Relationship

      Temporal Graph Attention

      Interactions become a continuous-time dynamic graph. Harmonic time encodings make reply cadence a first-class feature, surfacing topological disruption and authority inversion as the thread evolves.

      Intent

      Quantized Intent SLM

      A 1–3B model reads what is being asked: bank-detail changes, urgency, secrecy, approval bypass. It supplies the intent signal the graph cannot see.

      Boundary

      In-tenant, zero retention

      Inference runs inside your tenant or on-prem. Content is scored and discarded: no third-party cloud LLM in the path, nothing to retain.

      The research newsletter

      Teardowns, benchmarks and the math.

      High-signal technical write-ups for security leaders. The first issue lands in December.

      Get Issue #1 in your inbox

      Live attack teardowns, benchmark results and the architecture behind continuous-time graph defense.

      December v0.1 pilot

      Stop the thread, not just the message.

      Early access for Microsoft 365 enterprises and MSSPs. Tell us about your environment and we’ll be in touch.