Shadow AI and employee use of generative AI tools: how CISOs govern it

The sources and who is speaking

This explainer draws on three documents. Each one comes from a security leader or carries a security leader's words.

The first is a short blog summary on the Calian website. It points to an article by the company's corporate cybersecurity director, published on the CIO Association of Canada website.

The second is a webinar page from Obsidian Security. It features two speakers with CISO backgrounds.

The third is a Cornerstone OnDemand press release about an AI management certification, with a quote from the company's chief information security officer. The page header dates the release January 6, 2026. The dateline inside the text reads January 6, 2025.

None of the three is a full technical report. Taken together, they show the risk categories security leaders name and the control approaches they describe.

The risks named: data, privacy, access and side effects

The Calian piece sets out the problem in broad terms. It treats AI as useful but carrying its own hazards.

The piece then names four groups of issues for organizations to manage.

Each group maps to a question a security team can put to any generative AI tool that employees use:

• Data retention and ownership. What happens to the text, files or code an employee enters, and who owns it afterward?
• Transparency and privacy. Is it clear how the tool handles personal or confidential information?
• Access control. Who and what can reach the tool, and what can the tool reach in return?
• Unintended consequences. What outcomes follow that nobody planned for?

The Calian summary's title asks whether ChatGPT could compromise security. The linked article's address refers to AI applications exposing proprietary information. The summary itself does not list specific incidents or tools.

The Calian summary ties the answer to safeguards and the protection of sensitive information.

It also places responsibility on security teams to act early rather than wait.

Shadow AI: finding the tools people already use

The Obsidian webinar page uses the term shadow AI. It separates tools the organization has approved from tools employees adopt on their own. The first learning goal it lists is to find both kinds.

This is the point where employee use of ChatGPT-style tools becomes a governance problem. A security team can only apply the controls the Calian piece calls for to tools it knows about. The webinar page treats discovery as the first step, before any policy is written.

The page frames the overall aim as practical governance across the whole business, not just inside IT.

Identities, extensions and the browser

The Obsidian page goes beyond the chat window. It points to the identities that connect AI tools to company systems, including identities that are not people.

This links back to the access control concern in the Calian piece. An AI tool connected to a SaaS application, or acting through an account, can reach data well beyond what one employee types into a prompt.

The webinar summary also names browser extensions and integrations as areas to watch. It names visibility and browser-level controls as the way to protect sensitive data.

The page's site menus list "GenAI Data Leakage," "Shadow AI" and "AI Prompt Security" as use cases. That shows how the vendor sorts these problems. The page gives no detail on how its own products handle them.

Writing policy without stopping the work

The third learning goal on the Obsidian page concerns policy. It covers both generative AI tools and AI agents. It sets a condition: policy should not slow employees down.

The Calian summary makes a similar point. Both sources accept that staff will use these tools. Both present controls as the way to keep that use safe rather than as a reason to ban it.

From these two documents, the sequence CISOs describe is:

1. Find sanctioned and shadow AI tools.
2. Review the human and non-human identities that have access to sensitive data and key SaaS apps.
3. Watch extensions and integrations, and use visibility and browser-level controls to protect data.
4. Set policies for generative AI tools and agents that do not block productive work.

Formal governance: the ISO/IEC 42001 example

The Cornerstone release shows a more formal route. It reports a certification against a management system standard for AI.

The release describes what the standard covers.

It also lists the specific challenges the standard addresses.

In the release, the company's CISO links AI governance to the security role.

Dhanwantri describes what the standard gives the company in practical terms.

This certification concerns AI built into Cornerstone's own products. It is not about employees using outside chatbots. Still, the elements Dhanwantri names are clear policies, objectives and processes. Those are the same building blocks the Obsidian page calls for when it talks about setting policy for generative AI tools.

Why buyers and boards ask about AI governance

The Cornerstone release also quotes an outside researcher on how buyers now view AI governance.

The same researcher, Stacey Harris of Sapient Insights Group, gives a figure from her firm's research.

For a CISO, this places the governance work described above in a wider frame. The concerns about data, access and policy that the Calian and Obsidian documents raise for employee tool use are the same ones the research quoted by Cornerstone identifies as a barrier to adopting AI at all.

What the documents leave open

The three sources agree on the main categories:

• data handling and ownership
• privacy
• access
• visibility into which tools are in use
• clear policy

They give few specifics. None sets out a sample acceptable-use policy, a list of approved or blocked tools, or incident figures for data entered into chatbots. A security leader building a program would need other material for that level of detail.