Skip to main content

Coaching — Shadow AI Usage

Why it's risky​

Shadow AI is the use of unsanctioned AI services for work — typically free consumer tiers of public LLMs (ChatGPT, Claude, Gemini, Mistral, Perplexity), AI plugins in browsers, or "AI assistant" features in non-vetted SaaS. The 2024-2025 wave of "ChatGPT in every product" makes this nearly invisible.

The risk is not the model — it's the lack of contractual protection:

  1. No data processing agreement (DPA) — without a DPA, you have no contractual guarantee on data residency, retention, or use for training. GDPR, HIPAA and most enterprise compliance frameworks require one.
  2. No audit trail — sanctioned tools log usage to your SIEM. Shadow AI bypasses your DLP, your KMS, your IAM. You don't know what was sent or by whom.
  3. No customer-disclosure compliance — your contract with your customer may include "we will not share your data with sub-processors not listed in our DPA". Shadow AI silently violates this.
  4. No model governance — sanctioned models are evaluated for bias, safety, jurisdiction. Shadow AI is whatever the user found on Google.

The Noxys extension exists primarily to detect, classify, and coach away from shadow AI — converting it into either sanctioned use or a productive blocked moment.

Examples​

The user opens chatgpt.com from their corporate browser to get a quick
draft summary, while the corporate Copilot license sits unused.
The user installs a free "AI rewriter" Chrome extension that proxies
content to an unknown backend.
The user uses a "PDF summariser" SaaS that uploads internal documents to
its servers without the user realising it routes to OpenAI's free tier.

Best practice​

  • Know your sanctioned stack — your tenant has approved AI tools (Microsoft 365 Copilot, Azure OpenAI, internal Mistral, contracted Anthropic, etc.). Use them.
  • Check the catalog before using a new tool — the Noxys console exposes a catalog with sanctioned / under review / blocked per service. If the tool is not in the catalog, request a review.
  • Prefer enterprise-tier endpoints — even within the same provider (OpenAI), the consumer endpoint and the enterprise endpoint are different products with different contracts.
  • Report new tools — if you find a tool that solves your problem, request approval. The fastest way to convert shadow AI into sanctioned AI is to surface it to security.

Alternative phrasing​

❌ Risky workflow:

"I'll just paste this into ChatGPT free — it's faster than asking IT for the Copilot license."

✅ Safe workflow:

"I need a Copilot license for this kind of work. While I'm waiting, I'll rephrase the question to remove anything sensitive and use a public LLM only for the abstract version."

References​