Adverserial AI Enter platform

Confidential AI for cybersecurity

We built confidential inference.
Now we’re building confidential AI agents.

We make the full stack for security AI: our own cyber models, CyberKimi and CyberGLM, running inside hardware you can verify, and agents whose brokered actions are scoped, checked and recorded.

Explore the enterprise demo Try CyberGLM
  • 01Cyber modelsCyberKimi and CyberGLMAvailable
  • 02Confidential inferenceAttested, verifiable executionAvailable
  • 03Confidential agentsACK: scoped, checked, recordedMVP
How a confidential agent run works Illustrative workflow: your team sends an encrypted prompt to confidential inference, and a profile-bound agent uses capability-scoped tools. Each brokered action the agent proposes passes a capability policy that allows, denies or holds it for approval before it reaches Git, SIEM, EDR or cloud systems, and each decision is written to a signed trace. TRUST BOUNDARY / CONFIDENTIAL AGENT RUN ILLUSTRATIVE WORKFLOW YOUR TEAMPrompt +evidenceENCRYPTED ON DEVICE CONFIDENTIAL INFERENCE CyberKimiCYBER MODEL CyberGLMCYBER MODEL Confidential agentACK · SCOPED PROFILE INTEL TDX · NVIDIA H200 / B300 POLICY GitSOURCE CONTROL SIEMSOC TOOLING EDRENDPOINT CloudCLOUD WORKLOADS ALLOWED DENIED HOLD · APPROVAL SIGNED TRACEBROKERED DECISIONS, VERIFIABLE git.read · 7f3a siem.query · 1d08 cloud.delete · 9c41
Illustrative runPrompt → attested runtime → policy → tools → signed trace
01 / THE STACKFROM PRIVATE INFERENCE TO GOVERNED ACTION

One foundation.
Five connected layers.

From private model execution to governed enterprise actions. Each layer builds on the one below it.

The Adverserial AI stack: confidential inference, cyber models, confidential agents, capability enforcement and enterprise integrations, built upward from the foundation. 01 / CONFIDENTIAL INFERENCE KIMIGLM02 / CYBER MODELS ACK03 / CONFIDENTIAL AGENTS 04 / CAPABILITY ENFORCEMENT GITSIEMEDRCLOUD05 / ENTERPRISE INTEGRATIONS 01
01 / Confidential inferenceArchitecture
  1. 01 / The foundationAvailable

    Confidential
    inference.

    Run model requests inside an attested trusted execution environment. Verify the workload before sharing sensitive context.

    ProtectsModel inputs, outputs and execution
    Explore confidential inference
  2. 02 / Cyber reasoningAvailable

    CyberKimi
    and CyberGLM.

    Specialized cyber models for investigations, detection engineering and adversarial research. Red and blue team workflows on a confidential model route.

    ConnectsSecurity context to cyber-capable reasoning
    Explore the models
  3. 03 / Agent executionACK MVP

    Confidential
    agents.

    Bind each job to a profile, approved tools and scoped resources. Record the agent’s action path so an auditor can inspect its evidence.

    GovernsAgent identity, tools and resource boundaries
    Explore agent profiles
  4. 04 / Before the actionACK MVP

    Capability
    enforcement.

    Check each brokered action before execution. Deny unlisted capabilities, limit credential scope and hold sensitive actions for approval.

    ProducesPolicy decisions and signed action evidence
    Inspect a recorded agent run
  5. 05 / The enterprise boundaryIn development

    Enterprise
    integrations.

    Extend capability controls into Git, SIEM, EDR and cloud workflows through scoped connectors. Limit the action, destination and authority of each request.

    Extends toSource control, SOC tooling and enterprise systems
    Explore the integration demo

Private inference is the foundation.
Governed action is the next layer.

Explore the enterprise demo Confidential inference is available today. Profile-bound broker enforcement is demonstrated with Hermes and ACK. Enterprise connectors are in development.
02 / IN ACTIONACK + HERMES · DFIR INVESTIGATION

Watch a confidential
agent at work.

A DFIR investigation with ACK and Hermes, from policy setup to independent audit. Each brokered action is checked against policy and recorded for the auditor.

A DFIR investigation with ACK and Hermes, from policy setup to independent audit.

Explore ACK on GitHub
03 / OUR MODELSBUILT FOR CYBERSECURITY

Our own models.
Trained for security.

CyberKimi and CyberGLM are specialized cyber models, with red and blue team workflows on a confidential model route.

CYBERSECURITY MODEL / 01

CyberKimi

Deep context and operational clarity for investigations, detection engineering and adversarial research.

Parameters
2.78T
Active per token
104B
Context window
1M
Try CyberKimi
CYBERSECURITY MODEL / 02

CyberGLM

Security guardrails relaxed. Post-trained for red-team and blue-team operations.

Input
$2/ 1M
Output
$10/ 1M
For defenders
$0free
Try CyberGLM
BLUE TEAM

Detection engineering · Incident response · Threat hunting

RED TEAM · AUTHORIZED ENVIRONMENTS

Vulnerability research · Adversary emulation · Attack-surface review

CYBERSECURITY MODEL / 03  CyberSeek · A new addition to the cybersecurity model lineup.COMING SOON
04 / THE PROOFCYBERKIMI / PUBLISHED RUNS

Results,
not promises.

Open the CyberPVP evidence
CYBERPVP / CYBERGYM100-TASK COMPLETED RUN
97%

97 / 100 tasks solved
A completed, same-budget CyberPVP match. Three tasks reached a token or iteration limit.

SOLVEDNOT SOLVED
Inspect the sanitized traces
16/16

ExploitBench capabilities verifiedV8 · CVE-2024-6100View evidence package ↗

50runs

Recorded evaluation ladderThree-day campaign · score 2 → 16Read methodology ↗

ExploitBench assesses V8 d8; this result demonstrates renderer-equivalent V8 exploitation and does not claim a browser sandbox escape. Results published September 2026. Published Chrome/V8 research ↗

05 / FOR DEVELOPERSADVERSERIAL CONFIDENTIAL KIT

Meet ACK.
Code with
confidential inference.

One CLI connects your coding harness to CyberKimi and CyberGLM through verified confidential inference. Install ACK, enter your Adverserial API key, and launch the harness you already use.

  1. 01Install the Adverserial Confidential Kit
  2. 02Enter your API key in the masked prompt
  3. 03Choose an installed harness and start
Get ACK on GitHub Read the documentation
ACK / QUICKSTART
# 1. Install the current public ACK source
npm install -g \
  github:AdverserialAI/adverserial-confidential-kit

# 2. Enter your API key when prompted
ack init
# Input is masked. No key belongs in this command.

# 3. Inspect installed harnesses, then launch
ack harnesses
ack code hermes
VERIFIED MODEL ROUTECYBERKIMI + CYBERGLM
Selected harnessHermesChoose another integration below.

Nine harness integrations. One confidential model route.

Install the official harness CLI first. ACK configures its connection and launches an isolated, verified provider session.

ack code <harness>
Launch your selected harnessack code hermes

These integrations provide confidential model access. Profile-bound broker enforcement is currently demonstrated with Hermes and ACK; support for a model route does not imply the same tool controls across every harness.

06 / PRICINGSTART FREE · SCALE WHEN READY

Simple pricing.
Free for defenders.

Memberships with included weekly credit, or one prepaid wallet with no subscription. Pay only for the tokens you use.

MEMBERSHIPS$20/ month

Included weekly credit and higher plan limits, from Foothold to G0DMOD3.

Compare plans →
ENTERPRISETalkto us

Talk to us about your organization’s cybersecurity AI needs.

Contact us ↗
USD · MEMBERSHIPS BILLED MONTHLY · TOKEN RATES PER MILLIONSee full pricing
THE NEXT STEP IN YOUR SECURITY WORKFLOW

Put intelligence
into operation.

Specialized reasoning.
Built for the people who defend.

Explore the models
ADVERSERIAL AI
PROGRAM MEMBER
ENGINEERED FOR CYBER.

Before you continue

Clear responsibilities.

Use the models only for lawful, authorized activity. Review generated output before acting.

Terms & Acceptable Use · Privacy Policy · Terms PDF · Privacy PDF

This does not authorize a charge. Payment and subscription terms appear at checkout.