What Is Shadow AI? Risks, Examples, and How to Detect It

October 20, 2026 · 5 min read

Shadow AI is the use of AI tools — chatbots, coding assistants, browser-based generators, API playgrounds — by employees without the knowledge or approval of IT and security. It is the generative-AI version of shadow IT, and it is spreading faster, because the barrier to entry is a browser tab and a free account.

Why shadow AI happens

Nobody sets out to leak data. Employees reach for AI because it makes them faster: summarising a contract, debugging a stack trace, drafting a customer reply. When the approved tooling is slow, missing, or simply not there, people use what they already know. New AI products also appear weekly, so any fixed allow-list is out of date the day you publish it.

What shadow AI puts at risk

• Personal data. National ID numbers, customer details and HR records pasted into a prompt leave your control the moment you press Enter.
• Credentials and secrets. API keys and tokens end up in prompts when developers paste configuration or logs to get help.
• Source code and internal documents. Proprietary logic and roadmap material is shared with services you have no contract with.
• Compliance exposure. If you cannot say which AI services your people use, you cannot answer a regulator, an auditor or a customer questionnaire about them.

Why traditional tools miss it

Email DLP and CASB products were built for email attachments and sanctioned SaaS. AI traffic looks different: it is encrypted HTTPS to a long and growing list of domains, the sensitive content is free text inside a JSON request body, and much of it is sent from the browser or from developer tools such as editors and CLIs. A network firewall sees a connection to a domain; it does not see that the prompt contained a customer's ID number.

How to detect shadow AI

1. Observe on the endpoint. The most reliable vantage point is the device where the prompt is typed. An endpoint agent can route only AI-bound traffic through a local inspection proxy, leaving the rest of browsing untouched.
2. Separate known from unknown. Maintain a catalogue of sanctioned and recognised AI apps (for example ChatGPT, Claude, Gemini, Cursor, Perplexity) and treat any other AI-looking host as a discovery to review.
3. Measure before you block. Start by learning which tools are used, by how many devices, and how often, then decide what to allow, what to alert on and what to restrict.
4. Close the loop. Turn each discovery into a decision: approve the tool and add a policy, or retire it and tell the team why.

How Ladon helps

Ladon's endpoint agent routes AI traffic through a local proxy and reports which AI applications and unrecognised AI hosts are in use. The Shadow AI view lists each unknown host with hit counts, the number of distinct devices and when it was last seen, so you can see adoption as it happens rather than discover it in an incident review. From there you can set a policy per AI app: allow, alert, redact or block.

Shadow AI is not a people problem to be punished; it is a visibility problem to be solved. Once you can see usage, you can offer a sanctioned path that is as convenient as the one employees found on their own.