Architecture & Philosophy · Amplifier Safeguard

Trust at Scale

A tenant-level gateway that gives CIOs complete visibility and policy control over all AI agent activity in their organization.

Active Development · ramparte
May 2026 · ramparte/amplifier-safeguard
The Problem

CIOs can’t see it.
Can’t govern it.
Can’t audit it.

🕶️

Shadow AI

Thousands of employees using Copilot, ChatGPT, Claude, Cursor, and custom agents. No one knows what data is leaving the building.

🛡️

No Control Surface

Agents read documents, write code, access internal systems, and make decisions — with zero policy enforcement at the organizational level.

📊

Audit Black Hole

When compliance asks “what AI is doing in Legal?” the answer today is “we don’t know.” Every agent interaction is unlogged and invisible.

This is the “shadow IT” problem of the AI era — but with higher stakes. Agents don’t just access data, they act on it.

— Product Vision, amplifier-safeguard
The Insight

Sit inline.
See everything.

Palo Alto Networks’ foundational insight was that all network traffic flows through the network. Sit inline as a firewall — see everything. The agent equivalent: all agent work flows through LLM API calls.

From a single API call, the gateway can determine: WHO is using WHAT KIND OF AGENT with WHAT CAPABILITIES to do WHAT KIND OF WORK with WHAT DATA. No endpoint agent required.

— Architecture Design Doc, 02-ARCHITECTURE.md

Who

Identity mapping — user, department, role via gateway API keys linked to Entra ID

What Agent

System prompt reveals what the agent IS. Tool definitions reveal what it CAN DO.

What Data

Conversation history reveals what the agent HAS BEEN DOING — and with what data.

Architecture

Three layers, complete coverage

💬

Management Plane

Amplifier-powered NL interface — CIO talks to an agent, not a dashboard

🛡️

Gateway (Data Plane)

API proxy + Identity Mapper + Policy Engine + Event Store

🌐

DNS Backstop

Blocks direct LLM provider access via existing network infrastructure

Gateway handles

Sanctioned traffic — full inspection, policy enforcement, structured logging of every request/response pair

DNS Backstop catches

Shadow AI — personal API keys, consumer-facing tools, agents pointing directly at providers

Management Plane enables

Natural language policy creation, analytics, compliance — “Block Anthropic for Legal” just works

The Gateway

What the data plane sees

Every agent — Copilot, ChatGPT, Cursor, Amplifier, custom — must call an LLM API to think. The gateway intercepts that call and extracts three layers of intelligence.

1

Identity Resolution

Extract gateway API key, map to organizational identity — user, department, role, groups. Sub-5ms via in-memory cache.

2

Request Inspection

Extract provider, model, tool definitions, system prompt, conversation history, and content from the full API payload.

3

Policy Evaluation

Fast path (<10ms, deterministic): provider allow/deny, model restrictions, rate limits, regex patterns, tool restrictions. Slow path (async): LLM-based content classification, queued — never blocks inline.

4

Allow or Deny

Allowed: forward to real LLM provider, log everything. Denied: return structured error with reason, policy ID, and guidance. Total overhead: sub-50ms.

Policy Engine

Not just a telescope —
a control surface

Fast Path (Inline)

  • Provider allow/deny by department and role
  • Model restrictions per user group
  • Rate limits per user and department
  • Regex content patterns (SSN, credit cards, secrets)
  • Tool restrictions (no bash for non-engineering)
  • Time-based rules for contractors

<10ms evaluation · Deterministic · No LLM

Slow Path (Async)

  • Contextual PII detection beyond regex
  • Sensitive data classification (contracts, financials, source code)
  • Intent classification — what is the agent trying to do?
  • Prompt injection detection
  • Review queue for security team
  • Classification feedback loop

Never blocks inline · Retrospective · LLM-powered

Policies are created in natural language via the Management Plane. “Legal can only use Azure OpenAI” compiles into a structured policy, previews impact against historical data, then deploys with an optional grace period.

Threat Model

10 threats, one choke point

The agent IS the attack surface AND the worker. Safeguard detects threats at the LLM API bottleneck.

Prompt Injection

Adversarial input hijacks agent behavior via documents, emails, or web content

Data Exfiltration

Secrets, PII, trade secrets leaking through prompts to external LLM providers

Tool Abuse

A summarization agent that starts executing shell commands or accessing file systems

Privilege Escalation

Agent acquires credentials or tool access beyond intended scope

Supply Chain

Malicious MCP servers, tool packages, or agent bundles trusted and executed

Destructive Actions

rm -rf, git push --force to main, DROP TABLE — from hallucination or injection

Agent-to-Agent Poisoning

Compromised sub-agents returning malicious context to parent orchestrators

Compliance Violations

Unlogged access to regulated data — GDPR, HIPAA, SOX violations

Build Strategy

Five phases, each delivers value

1

Shadow AI Discovery (MVP)

API proxy + event store + NL analytics. The CIO gets their “I had no idea” moment. Value in just seeing what’s happening.

2

Policy Enforcement

Deterministic policy engine + NL policy creation + Entra ID integration. The gateway becomes a control surface, not just a telescope.

3

DNS Backstop

Config packages for Zscaler, Entra Internet Access, Palo Alto, corporate DNS. Full coverage — the gateway is no longer optional.

4

Content Intelligence

LLM-powered async content analysis, PII detection, sensitive data classification, review queue. Deep content awareness.

5

Endpoint Extension

MDM-pushed agent for local visibility — tool execution monitoring, agent inventory, unified timeline across API and endpoint.

Codebase & Velocity

~3 months of development

2.3K
Commits (non-merge)
41K
Lines of source
71K
Lines of tests
12
Source modules

Tech Stack

  • Python 3.12+ with FastAPI
  • uvicorn for async serving
  • httpx for outbound proxy calls
  • asyncpg for PostgreSQL event store
  • PyJWT + cryptography for identity
  • Pydantic for validation

Module Map

  • gateway/ — API proxy, caching, probes
  • policy/ — rule engine and compilation
  • identity/ — key-to-user mapping
  • events/ — structured event store
  • content/ — PII, classification, scanning
  • mgmt/ — 22 NL admin tools
  • dns/ — backstop config generators
  • endpoint/ — device-level agent
The Bigger Picture

Agentic
CASB

Cloud Access Security Brokers gave CIOs visibility into SaaS. Amplifier Safeguard does the same for AI agents — a new security category for a new era of compute.

What CISOs Get

  • Complete visibility — who, what, when, with what data
  • Policy enforcement without touching the agents
  • Full audit trail, searchable and exportable
  • Shadow AI discovery via DNS backstop
  • Compliance-ready content storage modes

Why It Works

  • Transparent to existing agents — no code changes
  • Multi-provider: OpenAI, Anthropic, Azure, Google, Mistral
  • Sub-50ms overhead on a 1–30 second API call
  • NL management — talk to an agent, not a dashboard
  • Uses existing network infra for enforcement

The insight is simple: all agent work flows through LLM API calls. Sit inline in that path, and you see everything agents think and do. No endpoint agent required. No framework integration required. No cooperation from the AI vendor required.

— Product Vision, 00-PRODUCT-VISION.md
Sources & Methodology

How this deck was built

All data sourced directly from the repository and git history. No metrics were fabricated or estimated.

Generated: May 2026 · Category: Architecture & Philosophy · Story angle: “Trust at Scale”

More Amplifier Stories