Your Employees Are Using Dozens of AI Tools. Do You Know What They Are Sending?

October 20, 2026 · 8 min read

A few years ago, "which SaaS tools does our company use?" was a hard question. Today it is a harder one: which AI tools do our people use — and what are they typing into them?

ChatGPT, Claude, Gemini, Perplexity, Cursor, coding copilots, AI note-takers, image generators, and a long tail of new products that launch every week. Your employees are adopting them because they work. The problem is that every prompt is a message leaving your organization, and in most companies nobody is looking at it.

The AI landscape your IT team never signed off on

Unlike traditional software, AI tools do not arrive through procurement. They arrive through a browser tab, a free account, or an editor extension. A developer pastes a stack trace. A support agent pastes a customer email. A finance analyst pastes a spreadsheet to "summarise the trends". Each one is trying to do their job faster — and each one has just sent company data to a third-party service you have no visibility into.

This is shadow AI: AI usage that happens outside the knowledge and control of IT and security. It is not malicious. It is simply invisible, which is exactly what makes it dangerous.

What exactly can go wrong?

Personal data leaves the building

National ID numbers, phone numbers, customer records and HR data pasted into a prompt are now outside your control. For organizations under privacy regulation — including Israel's Privacy Protection Law and its recent amendments — that creates real legal exposure, and it is very hard to answer a regulator's question if you cannot say what was sent where.

Credentials and secrets end up in prompts

Developers paste configuration files, logs and snippets to get help debugging. Those often contain API keys, tokens and passwords. A leaked key is not a theoretical risk; it is an incident waiting to be exploited.

Source code and trade secrets are shared

Publicly reported cases, such as engineers at a large electronics manufacturer pasting proprietary code into a public chatbot, show how easily this happens even in sophisticated companies. Intellectual property that took years to build can be shared in seconds.

You lose the ability to prove anything

Customers, auditors and boards increasingly ask: how do you govern AI use? Without inventory, policy and logs, the honest answer is "we don't know".

Why banning AI doesn't work

The instinctive response is to block ChatGPT at the firewall. It rarely works for long. Employees switch to a personal device or a different tool, productivity drops, and — worse — you lose the little visibility you had. Meanwhile your competitors are using AI to move faster. The goal is not to stop AI. It is to use it safely.

Why traditional security tools are blind to this

• Email DLP inspects email and attachments, not a JSON request to an AI API.
• Firewalls and web filters see an encrypted connection to a domain. They cannot see that the prompt contained an ID number or an API key.
• CASB and sanctioned-app lists only cover tools someone already knew about. New AI tools appear faster than lists can be updated.

Governing AI requires inspection at the point where the prompt is created: the employee's device.

How Ladon gives you control

Ladon is an AI governance and data loss prevention platform. A lightweight agent on each endpoint routes AI-bound traffic through a local inspection proxy — only AI services, so ordinary browsing is untouched — and applies the policies you define. In practice that means four capabilities:

1. Monitor — see every AI tool in use

Ladon recognises the major AI services (ChatGPT, Claude, Gemini, Cursor, Perplexity) and also flags unrecognised AI-looking hosts as shadow AI. The Shadow AI view shows each unknown host, how many devices touch it, how often, and when it was last seen. You get a real inventory instead of a guess.

2. Detect — find sensitive data inside prompts

Detection runs on the device. Ladon validates Israeli ID numbers by checksum and credit card numbers by the Luhn algorithm, and recognises common API key formats. Validation instead of naive pattern matching keeps false positives low, so alerts stay meaningful. For credentials you already know, a local secrets vault replaces exact values with an alias before the request leaves the machine.

3. Control — allow, alert, redact or block

Each policy ties an action to a data type and an AI app:

• Allow approved usage.
• Alert — let the request through, but record it and notify the team.
• Redact — replace the sensitive value with a placeholder and forward the rest, so the employee still gets their answer.
• Block — stop the request when there is no safe version of it.

Rules are evaluated by priority and the first match wins, so exceptions are easy to express: redact IDs everywhere, alert on secrets, and let one team use one tool freely.

4. Respond — incidents and alerts in real time

Flagged requests open incidents in a live console, with the AI app, the data type, the device and the action taken. A Slack webhook notifies your team as soon as a new incident opens. Everything is logged, so you can show an auditor exactly what happened.

Rolling it out without friction

1. Deploy silently through your MDM. Ladon includes a ready script for JumpCloud and a package for other MDM tools, enrolled with a single token.
2. Start in alert mode. Learn how AI is really used for a week or two.
3. Turn on redaction for IDs, cards and secrets.
4. Block selectively where redaction cannot make a request safe.
5. Tell your employees. Publish a clear monitoring notice — Ladon's Compliance page includes templates to start from (have your legal counsel review them).

Frequently asked questions

What is shadow AI?

Shadow AI is the use of AI tools by employees without the approval or visibility of IT and security teams.

Can I monitor what employees send to ChatGPT?

Yes. With an endpoint agent inspecting AI-bound traffic you can see which AI apps are used and flag requests that contain sensitive data, according to policies you set.

Can I block sensitive data without blocking ChatGPT entirely?

Yes. Redaction removes only the sensitive value and lets the rest of the prompt through, so people can keep using AI productively.

Does Ladon slow down normal browsing?

Only AI services are routed through the inspection proxy. Other traffic goes directly to its destination.

The bottom line

AI is not going away, and your employees are not going to stop using it. The choice is between blind adoption and governed adoption. With visibility, policy and enforcement in place, your organization can say yes to AI — without wondering what left through the back door.