How AIEGIS Works — Privacy By Design

Employee Laptop sends encrypted log stream to Customer's Cloud / VPC. Endpoint runs network proxy (detects AI API calls, captures prompt+response+tool calls). Browser extension catches Chrome/Firefox chatbot UI usage. Customer Cloud runs aiegis Dashboard (Docker) with search, real-time alerts, audit log export, policy engine. AIEGIS LIMITED cloud sees only license validity + software updates. NOT prompts, responses, customer data.

Captures: outbound network requests to AI vendor APIs.

Vendor coverage (browser extension, 20+ AI tools): ChatGPT, Claude, Microsoft Copilot, Gemini, Mistral, Perplexity, Grok, DeepSeek, Meta AI, HuggingFace, Poe, You.com and more. Cursor and VSCode Copilot via native app monitoring.

Per-call data: prompt, response, model, tool calls, latency, status, employee ID.

Does NOT: capture personal user content, decrypt non-AI traffic, run as root, modify packets.

Footprint (design target): <2% CPU, <50MB RAM, <100kbps bandwidth. v1 measurements published in pilot agreements.

Platforms: macOS (Endpoint Security framework + Network Extension) — available now; Windows (Filtering Platform driver) and Linux (eBPF, no kernel module) — coming soon.

Captures: prompts + responses inside chat interfaces of chat.openai.com, claude.ai, copilot.microsoft.com, gemini.google.com.

Why separate: web AI uses TLS to vendor cloud — endpoint sensor sees TLS handshake but not plaintext. Browser extension sees plaintext at DOM level.

Privacy-first: only runs on configured AI domains. Browsing other sites never inspected.

Deployed where customer chooses:

• On-premises — Docker / Kubernetes in customer's data center

Stack: PostgreSQL + Python backend + React UI. Open-source components, audit-friendly.

Features:

• Live feed of AI activity per employee
• Search by employee, model, data classification
• Real-time alerts on policy violations (PII pasted into ChatGPT, prompt injection patterns)
• Audit-grade export (CSV, PDF, JSON) — structured log export for evidence collection and regulatory review
• Role-based access (CISO sees all, manager sees their team only)
• Optional integration with SIEM (Splunk, Sentinel, Crowdstrike Falcon)

Data ownership: every byte in customer infra. AIEGIS LIMITED has no copy, no access, no key.

Minimum required to maintain the service:

• Software version / update status
• License key validity
• Aggregate health metrics (alive yes/no, last heartbeat)
• Anonymous error reports (no customer data, no employee identifiers)

That's it. We sell software. Customer owns data.

Potential supporting capabilities — applicability must be assessed per use case. Articles 9-15 impose requirements on providers of high-risk AI systems; whether any obligation is met depends on your system's classification and deployment context.

Article	What it requires	Potential capability (applicability assessed per use case)
9	Risk management	May support activities related to per-AI-call risk classification
10	Data governance	May support activities related to data-flow traceability
11	Technical documentation	May support activities related to system architecture documentation
12	Record-keeping	May support activities related to per-employee per-call logging
13	Transparency	May support activities related to disclosure-ready audit reporting
14	Human oversight	May support activities related to real-time policy enforcement
15	Robustness	May support activities related to anomaly detection on AI behavior
50	Transparency obligations	May support activities related to AI-generated content detection
72	Post-market monitoring	May support activities related to continuous behavioral baselining

Article 99 penalties range up to €35M or 7% of worldwide turnover for prohibited practices (Art. 5), €15M or 3% for other obligations, €7.5M or 1.5% for incorrect information — ceilings, not automatic. Article 50 transparency enforcement: August 2 2026. Articles 9-15, 72 (high-risk): December 2 2027.

We're transparent about scope. v1 covers approximately 80% of typical enterprise AI usage. The remaining 20% is on the v2 roadmap (3-6 months post-pilot):

• Plugin / tool-call telemetry — ChatGPT plugins, Claude tools, Cursor agents calling external tools. v1 captures the LLM call but not always the tool-call payload.
• Embedding endpoints — text-embedding-ada, Voyage, Cohere embeddings. We currently catch chat completion APIs; embedding APIs need additional enumeration.
• Local LLM (no network egress) — Ollama, llama.cpp, LM Studio running on localhost. We're researching syscall + model-file detection for v2.
• Mobile apps — ChatGPT iOS/Android. Requires separate mobile MDM integration, planned for v2.
• Voice input — Apple Voice → ChatGPT voice mode. We see the API call but not always the spoken prompt content.
• IDE plugin telemetry — Copilot in VSCode, Cursor, Windsurf. Some captured by endpoint network proxy; deeper IDE integration v2.

If your CISO use-case requires any of these for v1, talk to us — we can scope a custom integration in design-partner agreements.