Pre-breach protection for the AI era

AI security
you can prove.

CyberArmor protects AI in both directions — hostile content gated on the way in, secrets and dangerous output caught on the way out, and audit-grade evidence for every decision across 17 compliance frameworks, including SEC, FINRA, and NYDFS 500. Built for regulated enterprises — and anyone who knows they need AI security they can prove. See it on your own laptop in 15 minutes.

Run the 15-minute local PoC Request a Design Partner Pilot
Pre-ingestion URL trust gate — 15-minute local PoC
Runtime control: redaction, routing, identity, and audit evidence
Built for security-led pilots and design partners
cyberarmor.ai / platform / runtime-control

Actor

Identified

Tenant, user, app, agent, provider, and model context

Policy

Enforced

Monitor, warn, block, route, limit, or redact by context

Evidence

Recorded

Decision trace for SOC, audit, legal, and leadership

BLOCKEDURL trust gate: zero-width promptware detected — AI agent fetch blocked2s ago
REDACTEDCredential removed before AI submission — browser prompt18s ago
ROUTEDProvider policy applied — approved model path selected1m ago
2-Minute Overview

The Missing Security Layer.
Explained in Two Minutes.

Why enterprise AI has a gap no existing tool covers — and how CyberArmor closes it before content reaches your AI.

17 compliance framework packs
21 PII & secret redaction classes
SDKs & RASP in 9 languages
6 SIEM connectors
Endpoint agents on 3 operating systems
119 ms live end-to-end verdict
15-minute local proof-of-concept
The Problem

Enterprise AI Is Moving Faster Than
Security Can Follow.

Security teams are asked to govern AI systems, agents, providers, and workflows that outrun traditional review cycles. The hard part is not seeing AI risk. It is controlling it in both directions — what reaches your AI and what leaves it — and proving the control worked.

97%

of organizations that suffered an AI-related security incident lacked proper AI access controls.

63%

had no AI governance policy in place when the incident happened.

SEC · FINRA · NYDFS

Cyber and supervision rules at regulated financial firms already reach AI activity — the obligation to control it and evidence it applies now.

Source: IBM Cost of a Data Breach 2025 →

AI Activity Spreads Across Uncontrolled Paths

Employees, developers, contractors, apps, and vendors use AI tools, APIs, assistants, and providers outside formal review. A list of tools is not control. Security teams need control points that act at the moment AI activity happens.

Sensitive Data Leaks Before Anyone Sees an Alert

Credentials, API keys, payment data, bank details, PII, and non-public information get pasted into generative AI before any ticket, alert, or review workflow fires. Detection has to end in redaction, blocking, and evidence — not a dashboard entry.

Governance Without Enforcement Is Just a Document

Most AI governance starts as policy documents, committee decisions, and vendor questionnaires. Until those controls run at runtime and write decision-level evidence, every violation is an exception without response, proof, or accountability.

Why Existing Tools Miss This

Traditional URL filters were built
for human browsing.

A page can look completely harmless to a user while hiding instructions in CSS, comments, metadata, Unicode tags, or zero-width characters. Existing Safe Browsing, SmartScreen, and VirusTotal feeds answer “is this site malicious for a human?” — not “is this content safe for an AI agent to ingest?”

CSS and off-screen hidden text

display:none, visibility:hidden, opacity:0, font-size:0. Invisible to a user, read verbatim by an LLM.

Unicode-tag and zero-width encoding

Instructions encoded in Unicode tag characters (U+E0000–E007F) or zero-width spaces. Browsers render nothing; AI contexts ingest everything.

Metadata, comments, and structured data

JSON-LD, Open Graph tags, HTML comments, and schema markup are never shown to a human visitor. They can carry arbitrary instructions for an AI reader.

CyberArmor evaluates external content before it enters AI context, then allows, warns, redacts, sandboxes, blocks, or isolates based on tenant policy — with evidence written on every non-cached decision.

The Platform

An AI Security Runtime for Control and Proof.

CyberArmor.AI gates external content before your AI trusts it, and checks what your AI sends back out — redacting secrets and catching dangerous output before it reaches downstream systems. Detection, policy, routing, identity, response, audit, and evidence in one operating model, with every decision recorded.

01

Identify

Understand the actor, tenant, app, agent, provider, model, and data context behind an AI action

02

Gate

Evaluate external URLs and web content for phishing, promptware, and prompt injection before it reaches any human, browser, or AI agent

03

Inspect

Evaluate prompt risk, sensitive data, credentials, provider posture, and policy-relevant signals

04

Control

Monitor, warn, block, route, limit, or redact — on content coming in and on responses going out

05

Prove

Capture decision-level evidence that explains what happened, what changed, and which policy applied

Cross-Layer AI Security Context

Users & Identities
Models & Providers
Applications & APIs
Agents & Workloads
Data & Prompts
Evidence & Audit Trails
Explore the Full Platform
What Exists Today

URL Trust Gate runs end-to-end.

The 15-minute local PoC installer brings up the full gate stack on any developer laptop and submits four crafted attack pages — benign, CSS-hidden promptware, zero-width injection, credential-harvest — all producing live verdicts in under 120 ms.

Implemented and Tested

  • FastAPI service — end-to-end, health-checked, Prometheus metrics on port 8014
  • URL canonicalization, querystring redaction, homoglyph / punycode normalization
  • SSRF-guarded safe crawler — no cookies, no credentials, redirect-hop revalidation
  • Playwright detonation sandbox on an isolated Docker network (port 8015, internal)
  • Heuristic detection: prompt injection, credential harvest, brand impersonation, zero-width stripping
  • ML-based detection via detection service — phishing, promptware, DLP, IOC scoring
  • Google Safe Browsing v4, Microsoft SmartScreen, and VirusTotal v3 reputation feeds
  • Tenant allow/block lists via policy service (GET /policies?tenant_id=…&scope=url-trust-gate)
  • Policy decisions: allow, warn, redact, sandbox, block, isolate
  • Evidence writes to audit service on every non-cached decision
  • Consumer hooks: browser extension, endpoint agent, RASP Python, LangChain SDK, LlamaIndex SDK
See the full capability status Request a Design Partner Pilot

Flagship Proof — ROS 2 Robotics

We don't just flag prompts. We stop actuators.

A 9.0 m/s velocity command against a 2.0 m/s policy reached the wire as 2.0 m/s — and 0.0 under emergency stop. Verbatim transcripts from the Raspberry Pi 5 validation session.

Read the proof
Outbound Control & Evidence

What Leaves Is as Controlled as What Enters.

Gating hostile content before it reaches the AI is half the loop. CyberArmor closes the other half on the way out: secrets and PII are redacted before a prompt leaves, AI responses are checked for dangerous output before they reach downstream systems, and every decision is written to evidence. Not only input — output too.

Outbound — Data Protection

Redact the secret before it becomes an incident.

Credential leakage into generative AI is the leak nobody sees. CyberArmor inspects AI-bound content and redacts secrets, PII, PCI, NACHA, NPI, and non-public indicators before the prompt leaves — and records the control that ran.

The response coming back is checked too — insecure output handling, OWASP LLM02. Caught on the way out:

Shell command in an AI response
Script injection (XSS) in generated output
Browser-data exfiltration pattern

BEFORE

Summarize this log: OPENAI_API_KEY=sk-... and password=hunter22

AFTER

Summarize this log: OPENAI_API_KEY=[REDACTED-OPENAI-KEY] and [REDACTED-PASSWORD]

Evidence: decision=ALLOW_WITH_REDACTION, policy=redact-secrets, findings=OPENAI_API_KEY/PASSWORD, raw secret preview suppressed.

Evidence — Bound to the Control

Not just logs. Control you can prove.

Every gate, redaction, block, and route writes a structured record: which user or agent acted, which provider and model, what data classification was involved, which policy applied, and what response ran. The proof is attached to the action — which is exactly what an auditor, an examiner, or your own incident-response team asks for.

Timestamped Actions
Causality Chains
Policy Attribution
Audit-Ready Records
EVIDENCE STREAM — ILLUSTRATIVE SAMPLE
URL trust gate: BLOCKEDEVT-00483

Zero-width prompt injection in an external page — blocked before it reached AI context

POL-URL-01: Block Promptware2026-04-11 09:14:38 UTC
Outbound redaction: ALLOW_WITH_REDACTIONEVT-00482

API key removed from an AI-bound prompt before it left the browser

POL-DLP-04: Redact Secrets2026-04-11 09:14:32 UTC
Output check: BLOCKEDEVT-00481

Command-injection pattern detected in an AI response before downstream delivery

POL-OUT-02: Dangerous Output2026-04-11 09:14:28 UTC
Product Availability

A Clear Boundary Between What Is Deployable and What Is Being Expanded.

CyberArmor.AI is being built with security-led design partners and controlled pilot deployments. The platform already includes working control, detection, policy, redaction, routing, identity, audit, endpoint, and onboarding paths. Broader enterprise workflows are intentionally marked as pilot-stage while they mature.

Pilot-ready capabilities

Available Today

  • URL & Context Trust Gate — 15-minute local PoC with four live attack-page demos (display:none promptware, zero-width injection, credential harvest, benign)
  • SSRF-guarded safe crawler and optional Playwright detonation sandbox for JavaScript-heavy content
  • External reputation feeds: Safe Browsing v4, Microsoft SmartScreen, and VirusTotal v3 with in-process caching
  • Tenant URL allow/block lists with scope-based policy integration
  • AI request monitoring and policy decision logging
  • Prompt-risk, credential leak, and sensitive-data inspection
  • Redaction-mode policy decisions for supported browser, endpoint, SDK, extension, and provider paths
  • AI provider routing, provider resolution, credential-handling, and audit event patterns
  • Agent identity registration, tenant scoping, delegation metadata, and revocation paths
  • Endpoint-assisted AI tool and connection discovery
  • Tenant-scoped policy builder, artifacts, and API-key flows
  • Audit logs, telemetry, incidents, and evidence capture
  • Enterprise SSO (OIDC) with just-in-time provisioning and multi-factor authentication (TOTP + backup codes)
  • Directory identity enrichment — Entra ID, Okta, Ping, AWS IAM Identity Center — resolving every event to a named user, department, and group
  • 17-framework compliance policy packs including ISO 42001, SEC Cyber, FINRA Cyber, and NYDFS 500, with persisted tenant-scoped evidence
  • AI Bill of Materials (A-BOM) with OSV vulnerability scanning, enriched by CISA KEV and FIRST EPSS
  • Per-connector SIEM forwarding to Splunk, Sentinel, QRadar, Elastic, Google SecOps, and syslog/CEF
Expanding with customers

In Pilot / Design Partner Phase

  • Endpoint protection with automated, policy-gated patch remediation across Windows, macOS, and Linux
  • Signature and hash-reputation malware detection with a control-plane-synced intelligence feed
  • Expanded shadow AI inventory across SaaS and identity sources
  • Agent trust verification and delegation-chain workflows
  • Proof Vault evidence packaging and external audit exports
  • Advanced fail-closed runtime enforcement across more enterprise control points
  • Native PQC deployment validation where customer environments support required cryptographic dependencies
  • Deeper industry-specific compliance assessment depth with design partners

URL Trust Gate runs end-to-end in 15 minutes on a developer laptop — detection, policy, evidence, and four live attack-page verdicts.

Audit, action graph, compliance, and evidence paths are tied to enforcement decisions, not just reporting copy.

Endpoint agents, browser/IDE/Office extensions, SDKs, RASP hooks, and LangChain/LlamaIndex wrappers support pilot validation.

Built for pilots. Honest about the boundary.

Security buyers deserve to know exactly what runs in production, what is pilot-ready, and what is roadmap. Here is where CyberArmor stands today.

In production

The policy engine, detection (prompt injection, 21 redaction classes, dangerous output, toxicity), the compliance engine with 17 framework packs, persisted evidence, SSO with MFA, directory enrichment, and A-BOM vulnerability scanning run in the hosted stack and are supported.

Pilot-ready

Endpoint agents on Windows, macOS, and Linux, patch remediation, the URL Trust Gate with detonation sandbox, browser / VS Code / Office extensions, RASP and SDKs in nine languages, SIEM forwarding, and the ROS 2 robotics agent deploy today in controlled pilots.

Roadmap

The Windows kernel minifilter driver is not yet code-signed and does not ship. SOC 2 certification is planned, not attained. Self-service onboarding is still being refined.

See the full capability status
Compliance Evidence

Compliance evidence in the frameworks your examiners use.

Competitors treat compliance as a badge on the deck. CyberArmor ships it as a product: 17 framework policy packs run in the production compliance engine, and every assessment persists tenant-scoped evidence bound to the control decision that produced it.

FINRA 2026 Oversight Report

FINRA's 2026 Annual Regulatory Oversight Report expects member firms to log AI-agent actions and decisions, put guardrails on agent behavior, and monitor which systems agents can access. Those are the control decisions CyberArmor records — as evidence, bound to policy.

Shipped policy pack

SEC Cybersecurity

Maps control decisions to SEC cybersecurity risk-management and incident-disclosure expectations, with stored evidence for examination.

Shipped policy pack

FINRA Cybersecurity

Maps runtime controls — agent identity, action logging, guardrail decisions — to FINRA cybersecurity program expectations.

Shipped policy pack

NYDFS 500

Maps controls to the NYDFS Cybersecurity Regulation (23 NYCRR Part 500) for covered financial-services entities.

Shipped policy pack

SOC 2-Aligned Controls

Maps platform controls to SOC 2 Trust Services Criteria and stores the evidence. This is a control mapping — SOC 2 certification itself is on the roadmap, not attained.

Four of the 17 framework policy packs in the production compliance engine. The full list, with per-capability status, is on the capability status page.

Broader AI Governance Alignment

OWASP LLM Top 10Prompt injection, sensitive information disclosure, excessive agency.NIST AI RMFGovern, map, measure, and manage AI risk.ISO/IEC 42001AI management-system practices.EU AI ActDocumentation, logging, human oversight, robustness.
Use Cases

Real Threats. Real Buyers. Real Answers.

CyberArmor.AI is built around the security challenges enterprise teams are already facing: AI data leakage, prompt misuse, agent identity, provider sprawl, and the lack of proof when AI activity crosses a boundary.

CISO / Head of Security

Shadow AI Discovery & Governance

The Problem

Employees and vendors are connecting to AI tools, APIs, and models without security review. You have limited inventory, uneven controls, and little evidence when usage crosses policy.

The Solution

CyberArmor.AI uses supported endpoint, browser, API, and integration signals to surface AI usage, connect it to policy, and preserve evidence as coverage expands with each deployment path.

Explore Shadow AI Control
Why CyberArmor.AI

Five things that make this different.

01

Runtime control, not paperworkpolicy resolves to block, redact, route, or warn at the moment of the AI action.

02

Both directionshostile content is gated on the way in; secrets are redacted and dangerous output is caught on the way out.

03

Evidence bound to enforcementevery decision records the actor, the policy, the data classification, and the response that ran.

04

Content trust for AI agentsthe URL Trust Gate answers “is this safe for an AI to ingest?”, not just “is this site safe for a human?”

05

Honest boundariesproduction, pilot-ready, and roadmap are stated plainly on the capability status page.

Founder-Led

Built by a Security Practitioner for Teams That Need Control and Proof.

CyberArmor.AI is founder-led by Patrick Kelly — his full-time work — and built from the operating reality of enterprise security: policy has to become enforcement, sensitive data has to be protected before exposure, incidents need evidence, and AI adoption cannot wait for a perfect governance program.

The company is intentionally transparent about product maturity, design-partner work, and where the platform is strongest today. That posture is part of the product.

Read the company story

Security practitioner

Built from application, data, cloud, endpoint, identity, and AI security operating problems.

Hands-on builder

Rooted in working controls, tests, demos, runbooks, and deployment paths instead of slideware.

Enterprise lens

Designed for regulated environments, uneven ownership, legacy systems, and real security-team workflows.

Official brand and domains: CyberArmor.AI is operated by CyberArmor AI, Inc. Official web properties are served from cyberarmor.ai and its subdomains — app, admin, docs, and support. CyberArmor.AI is not affiliated with similarly named third-party domains, services, or social profiles unless they are linked from an official property.

FAQ

Questions Regulated Buyers and MSSPs Actually Ask.

How do we deploy — SaaS or self-hosted?+

Both. The hosted SaaS stack runs the production control plane — policy, detection, compliance, and audit services — and the same services deploy on your own infrastructure. The 15-minute local PoC stands the stack up on a single laptop, so a regulated firm proves the self-hosted path before procurement starts.

What does a pilot involve, and how long does it run?+

A pilot deploys the endpoint agent and the policy packs your firm is examined against, with success criteria agreed up front: policy enforced at your control points and evidence persisted for every decision. Plan for 30 to 60 days from kickoff to evidence review. Scope is set against the capability status page, so there is no ambiguity about what is production and what is pilot-ready.

Can our MSSP deliver and operate the platform?+

Yes. The platform is multi-tenant with tenant-scoped policy and evidence, so an MSSP operates multiple client environments from one control plane. SIEM forwarding delivers findings to Splunk, Sentinel, QRadar, Elastic, Google SecOps, or Syslog/CEF — into the SOC workflow the MSSP already runs. Select “MSSP / Channel Partnership” on the contact form to scope a delivery model.

What data does the platform see and store?+

Detection inspects AI-bound content at the control point and classifies prompt-injection risk, sensitive data and secrets across 21 redaction classes, dangerous output (command injection, XSS, browser-data exfiltration), and toxicity. Redaction removes secrets, credentials, and PII before content leaves the protected surface. Evidence records store labels, counts, policy identifiers, hashes, and action metadata — not raw secrets.

How does evidence map to SEC, FINRA, and NYDFS examinations?+

The compliance engine ships 17 framework policy packs, including SEC Cybersecurity, FINRA Cybersecurity, NYDFS 500, and SOC 2-aligned controls. Every assessment persists tenant-scoped evidence bound to the control decision that produced it, and per-framework assessment reports are scored and stored. When an examiner asks how a control operated, the answer is a record, not a recollection.

What is production today, and what is pilot-ready?+

Production: the policy engine, detection (prompt injection, 21 redaction classes, dangerous output, toxicity), the compliance engine with persisted evidence, SSO with MFA, directory enrichment, and A-BOM inventory with vulnerability scanning. Pilot-ready: endpoint agents on Linux, macOS, and Windows, patch remediation, the URL Trust Gate, browser / VS Code / Office extensions, RASP and SDKs in nine languages, and SIEM forwarding. The capability status page is the authoritative line between the two.

Does CyberArmor.AI replace our existing security stack?+

No. It integrates with the SIEM, SOAR, IAM, and cloud security tools you already run, adding the AI-specific layer they lack: actor and agent identity, model and provider usage, prompt-risk signals, policy decisions, and evidence records that forward into your existing pipeline.

What happens if the founder is unavailable?+

The platform was built end-to-end by its founding engineer — and the company around it is founder-led, with co-founder Alan Pan, an advisory team, and a go-to-market team. Continuity is engineered rather than assumed: deployment uses standard containers with documented runbooks, the full stack runs on your own infrastructure, and source-code escrow can be arranged as part of pilot or production contracting. Pilot agreements state these commitments in writing.

Get Started

Ready to Control and Prove AI Activity?
Let's Talk.

See how CyberArmor.AI maps to your AI activity, data leakage risk, agent workflows, provider usage, runtime controls, and evidence needs. The best demos start with the control problem you already have.

Scope a PilotRun the 15-Minute PoCContact the Founder

No spam. No hard sell. Every request is reviewed personally.

Before You Buy

See exactly what is production, what is pilot-ready, and what is not built yet →