▲AI GuardRequest a pilot
Workstation security

Credentials stay on the workstation. Protected when working with AI.

AI Guard intercepts terminal commands, MCP tool responses and clipboard — and masks credentials and PII before they reach the AI agent. Centralized policies enforce governance, and all processing runs on-premises.

On-premisesStrict modeLocal processing
The problem

AI tools read faster than DLP can react

A developer connects an AI agent — and it executes more commands in a minute than a human would in an hour. Every command returns output, and that output flows automatically into the external model's context: that is how agents work.

ai

AI agents in the terminal

AI agents execute commands and read output automatically. A single cat .env — and credentials are already in the external model's context.

>_

Secrets in command output

Logs with PII, configuration, infrastructure state and cloud credentials are routinely printed to stdout and flow downstream.

⧉

Clipboard

A copied token reaches an assistant chat or external web tool in seconds — faster than a network gateway can react.

⇄

MCP tools

MCP servers give agents direct access to files and resources. Local control closes potential leak channels at each tool boundary.

A network gateway sits further down the chain: data leaves the workstation before centralized systems can react. A local agent moves control to the point of origin.

How it works

Intercept — local check — audit

All processing happens on the workstation. The server receives only audit events, and the control plane is deployed in your infrastructure.

01

Agent intercepts

The local agent hooks into shells — zsh, bash, PowerShell — and intercepts commands and stdout/stderr streams via a secure PTY session, along with MCP traffic and clipboard.

02

Policies decide

A local rule-based engine evaluates each command and its output against signed corporate policies and applies a decision: allow, warn, mask or block — in real time.

03

Events stay on-prem

Every trigger is logged in a local audit journal and shipped in batches to the admin console inside your perimeter. Original credential values never leave the workstation.

AI security perimeter

Layered control of AI agent data and actions

AI Guard's architecture controls the full data and action path around the model: requests, commands, tool responses, local resources and audit events.

Layer

Comprehensive AI traffic protection

An LLM gateway inspects input and output between the user and the model. AI Guard extends that control to the terminal, MCP and clipboard used by AI agents.

Layer

Firewall logic for AI channels

A modern firewall applies layered defense: content inspection, action control, user context, access policies and event audit — all in a single execution chain.

Layer

Control before the model responds

The core principle: stop a leak at the point of origin, before credentials, PII or internal material enter the external model's context.

What the security layer does inside the developer perimeter

  • ✓Request inspection. Validate inbound commands, user prompts and tool-call parameters for credentials, PII, policy circumvention attempts and forbidden request classes.
  • ✓Response inspection. Check stdout/stderr, MCP responses and clipboard before they reach the agent or model: mask, redact or block.
  • ✓Action control. Policies define permitted commands, directories, hosts, MCP resources and operations per role and workstation.
  • ✓Context-aware policy. Decisions factor in the command line, user, group, environment, credential type, station mode and target tool.
  • ✓Built-in audit. The control plane receives trigger facts and metadata; original values stay on the workstation. This supports investigation within an isolated data perimeter.
Controlled threats
  • prompt injection via tool responses and documents
  • credential leaks from .env, logs, configs and stdout
  • excessive AI agent access to MCP tools
  • data exfiltration via clipboard and side channels

Practical takeaway: full coverage spans local commands, tool calls and responses, model requests and audit events in a single control chain.

Features

One agent — all leak channels

$

Terminal protection

Real-time streaming mask of stdout/stderr. Output stays readable while credentials are replaced with safe masks.

⇄

MCP protection

Local intermediary validates JSON-RPC requests, blocks access to forbidden resources and masks responses while keeping the protocol functional.

⧉

Clipboard protection

Background watcher finds secrets in copied text, replaces them with a mask and notifies the user.

⌕

Repository scanner

Local search for secrets and PII in the working copy: categorized reports, suppression rules, reproducible metrics.

§

Centralized policies

Draft → review → publish lifecycle, versioning, scope by users, groups and workstations, centralized distribution.

▦

Admin console

Workstation registry, health monitoring, security event log, policy editor, metrics and license management in a single interface.

Trust model

Strict mode: persistent workstation protection

  • ✓Signed policies. Every policy snapshot is signed by the server and verified by the agent locally.
  • ✓Integrity control. Manual modification of settings or policy is detected — the agent engages strict fallback mode and reports the event to the control plane.
  • ✓Persistent local protection. In strict mode the agent and shell integrations maintain active enforcement state.
  • ✓Minimal telemetry. Events contain trigger facts and metadata; original values stay on the workstation.

Honest guarantee boundary. Strict mode covers routine bypass, accidental leaks and one-click disablement. The threat model separately documents scenarios involving a user with full administrator privileges. Your team can test the agent during a pilot deployment and penetration testing.

Deployment

Fully on-premises

Server stack and data are deployed entirely within your infrastructure.

Server stack

Docker Compose, PostgreSQL and a reverse proxy. Health-check and readiness endpoints, structured logs, Loki export, documented backup and restore procedures.

Agent rollout

Packages for macOS and Linux; Windows and PowerShell targeted for Q4 2026. Deployment via MDM or software management, automatic shell integration and agent startup.

Scale

From a pilot stand to a fleet of 500–1,500 workstations. Documented resource sizing, scaling architecture and high-availability recommendations for the production perimeter.

Integrationsroadmap

SSO and AD groups with lifecycle management, event export to SIEM and Grafana with alerting, Kubernetes and Helm delivery, Guard mode for external agent frameworks.

docker composepostgresqlsigned policy versionszsh / bash / PowerShellmacOS / Linux
Who benefits

Value for every role

CISO / Security Lead

  • Centralized policies scoped to groups and workstations
  • All masking and blocking incidents in a single console
  • Verifiable integrity: signatures, tamper events, audit trail

Platform team

  • Stand up the perimeter using a documented Compose stack
  • Roll out agents through your existing MDM process
  • Diagnose via health-check endpoints and structured logs

Developer

  • Familiar terminal and tools — workflow stays intact
  • Local processing keeps the terminal responsive
  • Masking focuses enforcement on data while preserving developer privacy
Track record

Built on real-world deployment experience

The team behind AI Guard has delivered enterprise projects in telecom, fintech, adtech and consulting. Some engagements are covered by NDAs.

NDA-protected projects

Some AI and middleware infrastructure engagements are covered by non-disclosure agreements. Company names are listed publicly; implementation details remain within contractual boundaries.

  • AdSkill / AdWave LTD
  • Adsterra / AD Market Limited
  • Propeller Ads Ltd

Engagements delivered

  • LLM access and proxy infrastructure for AI model operations
  • API layer for AI and fintech workload integration and routing
  • Applied AI models for domain-specific use cases
  • Projects requiring access control, traceability and sensitive data handling
Team

Building AI security around the real development workflow

AI Guard is developed as an engineering security product: from the threat model and local policy enforcement to pilot deployments in customer infrastructure.

Founder

Tanya Farberg

Founder of AI Guard. Leads product, architecture and pilot engagements at the intersection of AI infrastructure, developer tools and workstation security.

Pilot conversations go directly to the founder, who designs the policy enforcement layer, governance model and guarantee boundaries.

Founder-led product

Architecture, threat model, pilot deployments and product decisions are driven by the team that designs and ships the system.

Developer tool security

Focused on real-world AI agent behavior across the terminal, MCP tools and developer workstations.

Pilot in your perimeter

Value is proven in your environment, on real repositories, against measurable acceptance criteria.

Trust and verifiability

Answers for enterprise technical due diligence

AI Guard is in the corporate pilot phase. Product status, architecture, metrics and guarantee boundaries are disclosed upfront.

Verifiable detection

  • Accuracy and recall measured on your repositories; metrics are reproducible
  • Thresholds and categories are configurable; suppression rules are part of the system
  • Pilot summary report includes measured indicators

Data and privacy

  • Data is deployed in your perimeter
  • Telemetry contains trigger categories and facts; original values stay on the workstation
  • Full control over data retention and access

Vendor continuity

  • The solution runs autonomously in your perimeter
  • Local perimeter stays operational regardless of external service availability
  • Source code escrow available on request through an independent agent
Roadmap

What's next: LLM Firewall at the AI Guard gateway

AI Guard evolves from workstation control to a full AI traffic security perimeter. The next stage is inspecting model requests and responses directly at the gateway.

Now

Workstation control

Terminal, MCP, clipboard and repositories governed by corporate policies. Secrets and PII masked before reaching the AI agent.

Next

LLM Firewall at the gateway

Inspect model requests and responses at the data-path entry point. Prompt injection, jailbreak and data leaks under control. Enabled for existing clients without integration changes.

Later

Red teaming and compliance

Automated attack-based testing, compliance reporting, SIEM and SOC integration.

Pilot deployment

Test on your repositories and your infrastructure

AI Guard is in the corporate pilot phase. We deploy the solution in your infrastructure and measure results together on a live repository.

  • Server stack deployed in your perimeter
  • Pilot group of workstations connected
  • Detection quality measured on your repositories
  • Acceptance criteria and guarantee boundaries documented before kickoff
  • Architecture diagram, resource sizing and deployment requirements
  • Operational independence: perimeter, data and agents stay under your control

On-premises architecture: pilot data stays in your infrastructure

or email us directly: tanya@ai-guard.pro

FAQ

Frequently asked questions

Where is data stored?

The server stack, database and all events are deployed in your infrastructure. Telemetry contains only trigger facts and data categories; original values stay on the workstation.

How does protection stay active?

In strict mode the agent, shell integrations and policies are protected against disabling and manual tampering. Modification attempts are detected, and the agent engages a fallback protection mode. Less critical groups can use a flexible mode with a local toggle.

Which operating systems and shells are supported?

Pilot perimeter: macOS and Linux with zsh and bash. PowerShell support and full Windows perimeter targeted for Q4 2026.

Will this slow down terminal work?

Processing runs locally using rules and regular expressions. Network calls are off the critical path, so the terminal stays responsive. Protection continues autonomously regardless of server availability.

How is the server stack deployed?

Docker Compose with PostgreSQL is deployed by your platform team using a documented runbook. We provide an architecture diagram, resource sizing and operational procedures including backup and restore.

How is this different from DLP or a secret scanner?

Traditional DLP controls the network and perimeter; scanners analyze code in the repository. AI Guard adds control at the point of execution — commands and their output on the workstation, where data enters AI tools. A repository scanner is included as a complementary layer.

How is product continuity ensured?

The solution is deployed entirely in your perimeter and operates autonomously. For critical deployments, source code escrow is available through an independent agent, released to the customer upon agreed conditions.

Where is data hosted?

The server stack and all data are deployed in your infrastructure. You maintain full control over data retention, access and jurisdiction.