Skip to content

Live redaction before you click enter.

100% Zero-Egress AI Security

Stop sensitive data leaks in ChatGPT, Copilot, Gemini, Claude, VS Code Copilot, native AI apps and CLI tools. Without the cloud compromise.

Real-time visibility, local threat detection, and zero-egress policy enforcement across every browser, desktop app, and analyst workflow.

The Threat Landscape

$4.88M

The average global cost of a single data breach.

Source: IBM Cost of a Data Breach Report 2024

The Architecture Shift

Inspection moves to the endpoint. Your prompts stay there.

Legacy AI security routes every prompt through someone else's cloud to decide whether it was safe to send. Guardrail Remote decides on the device that wrote it.

Legacy Cloud Scanners

proxy · SaaS
Endpoint Vendor cloud AI
  • Prompts leave the device for inspection. A second copy of your most sensitive data lands in a third-party security cloud.
  • Network round-trips on every send. Latency grows with distance to the nearest inspection point.
  • Keyword rules, binary blocks. Context-blind matching that frustrates users into Shadow AI.

Guardrail Remote

on-device
Endpoint + local engine AI
  • 100% on-device WebAssembly & ONNX engine. Pattern matching and semantic intent analysis run in the browser and on the desktop.
  • Ultra-low latency local inspection. Verdicts are computed in memory, with no network hop between the user and the decision.
  • Zero prompts sent to a third-party security cloud. Telemetry carries verdicts and metadata, never prompt text.

The Enterprise Suite

One engine. Every surface where prompts are written.

The same detection rules and the same on-device model, enforced identically wherever your teams use AI.

Browser

Browser Shield

Protects ChatGPT, Claude, and internal portals in real-time.

Desktop

Desktop Monitor Daemon

Local inspection proxy for VS Code Copilot, native AI apps, and CLI scripts.

Web

Text & Transcript Sanitizer

Paste logs, chat transcripts or documents and get a redacted copy back. Runs entirely in the browser, with no network calls and nothing stored.

Operational Modes

Two policy profiles. Set per user, or enforced fleet-wide by MDM.

The finance team and the SOC need different protection. Each profile tunes what the engine redacts, warns on, and restores.

Profile 01

Workforce Governance

For the general enterprise: safe daily AI use in every department, without breaking the workflows people depend on.

  • Financial DLP

    On-device intent analysis flags sensitive financial data and proprietary strategy, such as draft earnings or M&A plans, with a warn-on-send check that asks for a business justification.

  • Conversational guardrails

    Jailbreak and prompt-injection attempts are hard-blocked before the prompt leaves the browser.

  • Automatic redaction

    Secrets, credentials and PII are replaced with labelled placeholders as the user types or pastes.

Profile 02

Secure Analyst & Developer

For SOC analysts, DevOps and engineers who need AI to reason about real infrastructure.

  • Reversible tokenization

    IP addresses become consistent tokens like [IP_TOKEN_1] before sending, and the real values are restored in the model's reply.

  • Context stays intact

    The same address always maps to the same token, so the model can still trace and debug across logs and code without seeing the real values.

You type why does ssh to 10.42.7.18 time out?
AI sees why does ssh to [IP_TOKEN_1] time out?
You read Check the security group on 10.42.7.18 allows port 22.

Advanced Adversary Defense

Jailbreaks are stopped where they are typed.

  • Persona and override detection

    DAN personas, "developer mode" prompts and "ignore previous instructions" overrides are matched on the device, including spaced and leetspeak variants, and hard-blocked.

  • Local semantic scoring with MobileBERT

    A bundled MobileBERT model scores each prompt's intent on the endpoint, so novel phrasings that no signature covers still get caught.

  • Metadata-only telemetry

    Each verdict becomes an OCSF Security Finding with the pattern IDs, score and decision. The prompt text never leaves the device.

Threat response tiers

  1. Block

    A known attack pattern matched. The send is stopped and cannot be overridden.

  2. Warn model score ≥ 0.75

    Adversarial traits without a known pattern. The user sees a warning and may send with a justification.

  3. Log 0.40 – 0.74

    Sent without interruption and recorded as a suspected injection for your SOC.

Integrations & Compliance

Fits the stack your SOC already runs.

MDM fleet management

Fleet management via Microsoft Intune, Google Workspace, and Jamf (macOS browser support).

OCSF telemetry for Splunk and Datadog

Security Findings delivered by webhook, metadata only.

  • Microsoft IntuneMDM
  • Google WorkspaceMDM
  • JamfmacOS browsers
  • OCSFSecurity Finding
  • SplunkSIEM
  • DatadogSIEM
  • Chrome EnterpriseManaged policy

AI Compliance & Audit Readiness

Built for Enterprise AI Audits: SOC 2, HIPAA & PCI-DSS 4.0.1

Zero-egress local inspection gives auditors a control they can test: sensitive data is removed on the endpoint, and every decision is logged as evidence.

Privacy & Confidentiality

SOC 2 Type II

No prompt or data egress to Guardrail Remote or any external security cloud. Sensitive values are redacted on the device before a prompt reaches a third-party LLM vendor.

ePHI Protection

HIPAA

On-device redaction of electronic Protected Health Information, including medical record numbers, NPIs, health-plan IDs, SSNs and dates, before submission to web AI tools like ChatGPT or Claude.

Req 3 & 12

PCI-DSS 4.0.1

Local redaction of Primary Account Numbers (Luhn-validated) and cardholder data, including CVVs, expiry dates and track data, at the endpoint.

AI Governance

ISO/IEC 42001 & NIST AI RMF

Endpoint policy controls aligned with emerging AI governance frameworks: MDM-enforced policy modes, documented risk tiers, and an OCSF audit trail of every decision.

Evidence your auditors can sample

Every block, warning and redaction is an OCSF Security Finding in your SIEM, with no prompt text inside it.

Request the control mapping

Replace legacy gateways with a lightweight edge engine (starting at 26MB).

Or email enterprise@guardrailremote.com