Red team and blue team · one engine

Anyone can be talked into it. Now anything can.

Criminals clone a voice to move money and reset the wrong account. Now they hide instructions inside a website to steer someone's AI assistant. DeepBlocker runs the attack the way they would, shows you exactly what got through, and blocks it live. Offence finds the gap, defence closes it, and the evidence proves it to a board, an insurer or a regulator.

Tuned on real G.711 telephony · screening live calls in production today
SESSION 0x8F2A · INBOUND · G.711 / 8kHzANALYZING LIVE
CALLER  +44 1534 ••• 902
CLAIMS  "M. Dupont, authorising payment"
LINE    Payments desk · high-value
1 · Is the voice live & human?
Live human, no synthesis detected
2 · Is it the enrolled person?
Matches M. Dupont · score 0.94
3 · Does the request make sense?
Known counterparty · within policy
VerdictVerified, allowsealed · ~1 s
The scale

This is not emerging. It is here, and compounding.

Cloned voices and faces are already moving real money out of real organisations. The figures below come from named, public sources.

£1.17bn
stolen from UK fraud victims in a single year
UK Finance, 2025
$25.6M
wired in one deepfake video call, across 15 transfers
Hong Kong police / CNN, 2024
60%
of firms hit a deepfake-driven incident last year
Thales, 2026
704%
rise in face-swap injection attacks
iProov, 2024 to 2025
$16.6bn
reported losses to cybercrime in one year
FBI IC3, 2024
73%
of organisations hit by cyber-enabled fraud
WEF, 2026
Beyond the single call

Real attackers do not make one call. They make several.

A first call is friendly and forgettable. It picks up a name, a system, a schedule, nothing that raises an alarm. A second call follows up days later, sounds familiar, and earns one small yes. The third call spends that trust on the real ask: a code, a reset, a payment. This is close to how the 2023 Retool breach actually ran: a text message, then a phone call, then an escalating request for an MFA code. Attack Sequences now runs the same chain against your own team, on a schedule, so you find the gap before a real attacker does. You pick from a library of these chains, each modelled on a documented breach, and the caller can even carry what it learned from one campaign into the next, months later.

CALL 1 · FIRST CONTACT
A friendly recon call
The caller asks a few harmless questions and picks up a name, a system, a schedule. Nothing about it feels like an attack.
CALL 2 · DAYS LATER
A rapport call
The same caller follows up, sounds familiar, and asks for one small favour. The target agrees without thinking twice.
CALL 3 · THE ASK
The exploitation call
Armed with what it learned and the trust it banked, the caller asks for the code, the reset or the payment.
MEMORY
The caller remembers everything from the last call and never starts over, exactly like a real attacker working the same target over days or weeks.
PROOF
Every simulated call is then replayed through DeepBlocker’s own voice detector, so your report shows not just who was fooled, but how many of these calls a machine layer would have caught. You test your people and your defense in one run.
What you can buy

One engine. Two jobs. Two channels.

Every product does one of two jobs. Red team finds the gap by running the attack. Blue team closes it by blocking what gets through. Today that runs on your phone lines. Now it runs on the AI assistants your staff and customers let act on websites too.

Voice · the phone line

Where the attack lands today. Two protect you from calls coming in, one protects a phone company from the calls its customers send out, and one is the raw sensor for developers.

PRODUCT 01
Assess
RED TEAM

Vishing Readiness Assessment

FOR → Any organisation that wants to test its resilience to AI voice fraud
  • Our AI red-team agent calls your staff and attempts a live voice-cloning fraud
  • You get a scored report: who was fooled, where the process broke, how to fix it
  • Upgrade to quarterly re-testing with a versioned evidence pack for your board, your regulator and your insurer
Full product page →
£9k-22k
one-off assessment
£2.5k-4k /qtr
recurring evidence pack
● Available now
PRODUCT 02
Shield
BLUE TEAM

Zero-Trust Voice Agent

FOR → High-risk phone lines and sensitive conversations
  • An AI agent answers or monitors high-value lines and runs all three checks in real time
  • Verified Caller List: enroll your 30 to 200 authorised voices, everyone else is challenged
  • Low risk passes through; high risk is held, escalated, and logged with full context
Full product page →
£250-500
per protected line / month
● In pilot
PRODUCT 03
Network
CARRIER

Outbound Screening + Compliance Records

FOR → Telecoms providers and voice platforms
  • Screens the calls leaving your network, starting with automated and AI-assistant calls
  • It blocks fraudulent and unauthorised synthetic-voice calls before they reach the public phone system
  • It writes a locked record of every check, so the carrier can prove to the regulator it was screening
Full product page →
£25k-50k
pilot + integration
Platform + per-channel
ongoing, at scale
● In discussion
PRODUCT 04
Detect
API

Deepfake Detection API

FOR → Developers & voice-agent platforms (Vapi, LiveKit, IVR builders)
  • The raw liveness sensor as a metered API, one call, one score
  • Tuned for G.711 telephony, returns live / synthetic / uncertain
  • The building block. Refreshed every quarter as new voice AI ships
Full product page →
Metered
per minute analysed
● Early access
AI agents · the browser

Where it lands next. AI assistants now take real actions on websites inside a logged-in session, and they can be talked into the wrong one. Same engine, same evidence pack, a target that is not human.

How it works

One engine. Three checks. Every call.

Most voice detection relies on a single AI model. DeepBlocker doesn't. Every call goes through three independent checks before it is trusted, so if an AI-generated voice gets past one, the next catches it. Because we verify your real voice, not the latest AI model, protection stays accurate as voice technology evolves.

CHECK 01Liveness

Is the voice live and human?

Our deepfake detector, trained on degraded real-world telephony where lab models collapse. Catches synthetic and cloned speech.

STOPS → voice clones, TTS, replay
CHECK 02Verification

Is it the enrolled person?

Each authorised voice is enrolled once. Every future call is matched against that fingerprint. New voice AI does not change what your real people sound like.

STOPS → human impostors, unknown callers
CHECK 03Context

Does the request make sense?

An AI agent reads the conversation and scores the ask against policy: amount, counterparty, urgency, pressure tactics.

STOPS → social engineering, coercion
ZERO-TRUST
Deny by default, allow the verified. Every single call resolves to a clear green, amber, or red verdict, and every screened call is written to a sealed record.
Tuned on real telephony, not lab audio
94.7%
deepfake recall on real phone-line audio (G.711)
≈1 s
verdict on a live call
Version-locked
models, hash-locked evaluation sets
UNCERTAIN
honest band, routed to human review
The part nobody else sells

Accuracy fades. Evidence lasts.

Every voice-AI vendor sells you a number: "we're 96% accurate." A number you cannot verify, from a model that changes next month. When a regulator or insurer asks what happened on one specific call, a number is worthless.

DeepBlocker seals every verdict into a record you can prove: what ran, which model version, what it decided, when. Not "trust us." Verify it yourself.

Sealed logs
Every call cryptographically fingerprinted (SHA-256). Nobody can edit the record after the fact, including us.
Versioned
The exact detector version is stamped on each verdict, so your evidence still holds up months later.
Honest
When the model is unsure it says uncertain and escalates, it never fakes confidence.
Vendor-independent
The record is not tied to one detector model or one voice vendor. An assessment grades the controls you actually run, whoever built them, including ours, against the same sealed standard.
Call evidence recordsealed
session0x8F2A · 17E4
timestamp2026-07-04 14:22:07Z
detectordb-v7-opus · AUC 0.946
livenessHUMAN 0.97
speakerMATCH 0.94
contextIN-POLICY
verdictVERIFIED · ALLOW
seal
sha256:9f2ac41e7b3d…08e4d1a6
🔒  Tamper-evident · admissible in audit
Why now

Liability is moving onto the network.

Just as STIR/SHAKEN forced every US carrier to install caller-ID authentication, fraud-call screening is becoming mandatory plumbing. The rules being written now will fine the carrier, per call, unless it can prove it was screening. That proof is the product.

FEB 2024
The FCC rules AI-generated voices in robocalls illegal under the TCPA. The clock starts.
APR to MAY 2026
The FCC proposes a $2,500 per-call fine for carriers that let bad actors onto their network, and asks whether carriers running effective AI screening should get a safe harbor from those fines.
JUL 27, 2026 · OPEN NOW
The public comment window closes. The rules that follow will set how carriers must screen AI voice fraud, and what they have to prove they did. Organisations preparing now will be the ones ready when the duty lands.
LATE 2026 / 2027
The final US rule is expected. The FCC also floats independent auditor verification of compliance, the exact seat an independent assessor fills.
2027 to 2028 · EST. 50 to 60%
Ofcom and the EU are expected to follow with analogous carrier duties. EU AI Act Art. 50 disclosure rules are already enforceable from Aug 2026.
Start here

Find out who on your team a clone could fool.

A Vishing Readiness Assessment is the fastest way to see your exposure, and the first page of an evidence trail you will be glad to have.