Bareket AIPowerGuardBAREKET AIRequest a pilot
AI INFRASTRUCTURE SECURITY

The control plane
for the power layer.

The power layer beneath every AI server, finally watched. Attacks caught, crypto miners exposed, stolen keys spotted, in milliseconds.

01Power attackOne command at the power layer can crash a server or burn out the hardware. Caught in about 20 ms, 571 times out of 571.02Crypto miningHidden miners hijack AI servers and steal power and compute. PowerGuard exposes them by the power they draw.03Crypto keysSecret encryption keys leak through how much power a chip draws. PowerGuard spots the attempt at the rail.

Purpose-built for the infrastructure that cannot fail

AI DATA CENTERSTELECOMFINANCIAL SERVICESDEFENSEGOVERNMENT
AI DATA CENTERS

The densest racks ever built.

Every new gigawatt of AI capacity arrives with thousands of power controllers nothing in the security stack defends. PowerGuard adds the missing line of defence in software, on servers already racked.

Request a pilot
TELECOM & CRITICAL UPTIME

Unattended sites, zero maintenance windows.

Core and edge sites run for years without hands on the hardware. The same measurements that catch an attack also catch power components drifting toward failure, before they take a site down.

Request a pilot
DEFENSE & GOVERNMENT

Sovereign by design.

Runs fully inside your environment, including disconnected networks. Machine telemetry only: no workload content, no customer data, ever. Enforcement acts under policy you set and can revoke.

Request a pilot
THE THREAT IS REAL

The layer beneath the OS is under active attack.

The power-control layer is no longer theoretical exposure. Industry research and confirmed incidents show it being probed, breached, and abused now.

Third-party incidents and research. Evidence the threat is real, not claims PowerGuard detected these.

Tens of thousands of BMCs exposed to takeover.

Server management controllers reachable from the open internet, from flaws first flagged over a decade ago.

Black Hat USA 2026, runZero

All four telecom operators in Singapore breached.

State-linked group gained months-long persistent access using zero-days and rootkits.

CSIS, Feb 2026

Miners deployed within 10 minutes of access.

Cryptomining campaign used compromised credentials to deploy miners on cloud compute in ten minutes.

AWS GuardDuty, Nov 2025

From cryptomining to supply-chain compromise.

One actor escalated from mining to compromising trusted security tools and AI gateways, planting backdoors for lateral movement.

Palo Alto Unit 42, 2026

Critical GPU host takeover, CVSS 9.0.

NVIDIAScape: a three-line container escape grants full root on the GPU host, systemic across managed AI clouds.

Wiz, CVE-2025-23266

Remote silicon destruction, proven.

PMFault drove voltage out of spec over PMBus to permanently damage server CPUs, remotely, no physical access.

Black Hat Asia 2023, CVE-2022-43309

Every one of these ends at the same place: the power-control layer inside the server. The layer below your security stack, and the one PowerGuard was built to see.

ADVANCING SECURITY FOR THE AI SUPERCYCLE

Your security stack ends at software.
PowerGuard goes deeper.

A compromised administrator or firmware component can issue commands beneath SIEM and EDR visibility. PowerGuard creates an independent control plane for the physical power layer.

01

Discover

Every power chip, controller and rail on the servers you already run, found and mapped on day one.

See discovery
02

Observe

PowerGuard watches the power nonstop and tells normal AI load from a real attack, in milliseconds.

Explore detection
03

Enforce

Its verdict reaches the controls you already run, only on real evidence, under the policy you define.

See enforcement
ONE INDEPENDENT CONTROL PLANE

Infrastructure that can
detect, decide and defend.

PowerGuard turns invisible machine behaviour into threat verdicts that security and infrastructure teams can act on, while the evidence stays inside your environment.

  • Continuous attack-surface discovery
  • 100 Hz detection, on the server
  • Physics-based attack detection
  • Evidence-led enforcement
PowerGuard live security dashboard in Datadog: verdict state, power, anomaly score versus adaptive fence, ISOLATE events
Live PowerGuard dashboard in Datadog · running on our bench
HARDWARE-VALIDATED

Protection measured
in milliseconds.

Built from real hardware research with transparent validation boundaries, not a generic anomaly engine wrapped around power data.

20 ms

Live attack detection

Measured during a physical voltage attack.

100 Hz

Continuous detection

Per controller, on your hardware, no cloud round-trip.

1–5 servers

Pilot scope

4–6 weeks, success criteria agreed in advance.

PMFAULT / LIVE HARDWARE PROOF

5.21 V → 4.64 V.
Detected before damage.

A live command drove a physical rail outside its safe behavior. PowerGuard observed the evidence and produced a verdict in 20 milliseconds.

REAL HARDWARELIVE ATTACKREPEATABLE TEST
DEPLOY WITH CONTROL

Start by mapping the surface.
Earn the right to enforce.

01

Discover

Map the attack surface without installing anything.

02

Observe

Detect in record-only mode, every verdict logged, nothing acted on.

03

Enforce

Act only after evidence and policy are agreed.

AI infrastructure protecting AI infrastructure.

The AI is behavioural, not generative. It learns what your hardware normally does, and reports what does not fit.

START WITH 1–5 SERVERS

See the power layer.
Then decide how far to go.

A focused 4–6 week pilot on your own or production-identical hardware, with measurable success criteria agreed in advance.

Start a pilot conversation