In the Age of AI, What New Risks Do Enterprises Face When It Comes to Confidential Data?

From AI models and AI agents to enterprise documents, as AI begins to “work for us,” it is time to rethink how confidential data should be protected.

One Lobster Makes Us Rethink AI Security Risks

t may look funny, but this scenario is actually not that far from reality.

As generative AI and AI agents become part of everyday business workflows, AI is no longer limited to answering questions. It can increasingly read documents, organize information, write reports, and even interact with other tools.

This raises a new question for enterprises.

Instead of simply asking:

“Will an employee send confidential data outside the company?”

we also need to ask:

“What data is the enterprise actually allowing AI to access?”

When we talk about confidential data leakage, we traditionally think of USB drives, personal cloud storage, email, or messaging applications.

AI is making the situation more complicated.

Today’s AI systems are increasingly capable of reading data, using tools, executing tasks, and even working autonomously for extended periods of time.

AI Is No Longer Just a “Chatbot”

Recent AI development has been moving in two clear directions:

more capable models and more capable AI agents.

For example, OpenAI’s latest GPT-6 Astra has expanded capabilities in computer use, browsing, software engineering, science, and professional work. Anthropic’s Claude Opus 5.5, released in September 2026, also focuses on complex professional work and software development.

But from an enterprise security perspective, the development of AI agents may be even more significant.

In September 2026, OpenAI introduced the Agents API, allowing developers to build systems that can operate for extended periods, process files, execute code, use tools, and coordinate sub-agents.

Meta’s Muse, introduced around the same time, demonstrates another direction for AI agents. With user authorization, Muse can use a dedicated virtual environment and browser to perform tasks such as sending emails, planning trips, and handling more complex workflows.

This represents an important shift:

AI is moving from “help me answer” to “help me get things done.”

And as AI begins to perform more work on behalf of enterprises, another security question naturally emerges:

What data is AI allowed to access?

What New AI-Related Risks Could Expose Enterprise Confidential Data?

01 | Employees Voluntarily Give Confidential Data to AI

This is perhaps the most straightforward scenario.

An employee may simply want to ask an AI tool to:

• Analyze a customer list

• Summarize unpublished financial information

• Modify a product design document

• Analyze competitor information

• Draft an unpublished business proposal

• Review source code for potential problems

From the employee’s perspective, these may simply be ways to improve productivity.

From an enterprise security perspective, however, the key question is:

Is this data actually allowed to be provided to this AI service?

As organizations begin using multiple AI tools internally, relying solely on employees to make these decisions on their own can make data governance increasingly difficult.

02 | AI Agents Can Actively Access and Process Data

The data flow of a traditional chatbot is relatively straightforward:

User → Input → AI → Response

An AI agent, however, may follow a much more complex workflow:

User → AI → Search for data → Open documents → Use tools → Execute code → Generate results

As agents gain access to more tools and permissions, enterprises need to manage more than just what employees enter into AI systems.

They also need to consider:

What is the AI allowed to read?

What can the AI operate?

Where can the AI send data?

What actions can the AI perform on behalf of the user?

For this reason, adopting AI agents is no longer simply a software procurement decision. It is increasingly connected to access control and data governance.

03 | Once Data Enters an AI Platform, Enterprises Still Need to Understand How It Is Processed

This is an issue that can easily be overlooked.

Some enterprise AI services now provide extensive data protection mechanisms.

For example, OpenAI states that inputs and outputs from commercial services such as ChatGPT Business, Enterprise, and the API are not used to train models by default. Enterprises can also use access controls and data retention settings to manage their data.

But enterprises need to understand an important distinction:

“Not used to train the model” does not mean “the AI does not process the data.”

AI must receive and process the information provided by users in order to perform tasks such as summarization, analysis, translation, and code generation.

Therefore, enterprises need to ask more than:

“Will the AI provider use my data to train its models?”

They should also consider:

• How will the data be processed?

• How long will it be retained?

• Which services or features can access it?

• Who can use it?

• Will it be connected to other systems?

• Are appropriate access controls and management mechanisms in place?

These are fundamental data governance questions that enterprises need to consider when adopting AI.

A Recent Case: AI Research and the Question of Data Boundaries

In September 2026, a controversy surrounding mathematical research provided a concrete example of why these questions matter.

New York University mathematician Tristan Buckmaster and Anthropic mathematician Levent Alpöge had previously used AI tools, including OpenAI Codex and Claude, in their mathematical research.

OpenAI later published research results concerning the Navier–Stokes Millennium Prize problem, prompting Buckmaster to question whether unpublished research data could have influenced OpenAI’s work.

According to reports, Buckmaster was concerned about whether unpublished research materials entered by him and his research collaborators into Codex could somehow have influenced subsequent results.

However, an important distinction needs to be made:

There is currently no evidence that directly proves OpenAI obtained and used this private research data.

OpenAI subsequently stated that, following an investigation, prompts submitted by Buckmaster to Codex during the two months before the relevant announcement could not have influenced the system in any way, including through training. OpenAI also stated that the two proofs differed substantially.

Therefore, the key lesson for enterprises is not to jump to a conclusion about whether the data was actually used.

Rather, it is to ask:

When highly confidential, unpublished information enters an AI platform, to what extent can an enterprise or research organization verify and control the boundaries of how that information is used?

This question may become increasingly important as AI becomes more deeply integrated into professional workflows.

What Can Enterprises Do Today?

AI security is not simply about “banning employees from using AI.”

A more practical approach is to manage AI-related risks at several levels.

Level 1: Establish an AI Usage Policy

Enterprises can define:

• Which AI tools employees are allowed to use

• What types of data may be entered

• What types of data are prohibited

• Which departments can use specific AI capabilities

• What permissions AI agents can have

This provides the basic governance framework.

Level 2: Choose AI Services with Enterprise Data Protection Mechanisms

When adopting AI, enterprises should look beyond model capabilities and understand what data protection and management mechanisms the service provider offers.

For example, OpenAI’s enterprise offerings provide mechanisms for enterprise data controls, identity verification, access control, and data encryption.

Therefore, when selecting an AI solution, enterprises should ask more than:

“Which AI is better?”

They should also ask:

“Does this AI’s data handling approach meet our organization’s security and data governance requirements?” 

Level 3: Control Who Can Access Confidential Data

In addition to managing what data AI is allowed to access, enterprises need to consider an even more fundamental question:

If a document contains confidential information, who is authorized to access its contents?

Encrypting important documents and using access controls to restrict decryption and use allows enterprises to manage not only where data can go, but also who can access the data itself.

This is a different layer of protection from simply preventing employees from uploading documents to a particular AI platform.

The first focuses on:

“Who is authorized to access the data?”

The second focuses on:

“Where can the data be sent?”

As AI adoption continues to grow, both questions deserve careful consideration.

Even when an enterprise allows employees to use AI, that does not mean every confidential document should always exist in an unprotected form or be freely accessible to every user.

From AI Data Governance to Protecting the Files Themselves

Data protection in the AI era does not necessarily have to begin with the AI platform or network traffic.

Enterprises can also start with a more fundamental question:

If a document itself contains confidential information, should it be protected from the very beginning?

goPatrol uses file encryption as a core protection mechanism and can manage how different users access protected documents according to the organization’s permission settings. It can also be combined with file operation logging, external device management, web upload controls, and print controls to help reduce the risk of confidential documents being accessed or taken without authorization.

In other words, enterprises can consider two questions at the same time:

What data is AI allowed to access?

and:

Who is authorized to access that data?

The first is part of AI usage and data governance.

The second is an existing file security challenge that remains important in the AI era.

These two approaches do not replace one another. Instead, they can work at different layers to build a more comprehensive data protection strategy.

Conclusion | The More AI Can Do, the More Data Permissions Matter

In the past, enterprises often asked:

“Did an employee send a confidential file outside the company?”

In the AI era, we also need to ask:

“Was the AI actually authorized to access that confidential file?”

AI is evolving from a tool that simply answers questions into a digital agent capable of reading data, using tools, and completing tasks on behalf of people.

The real question is not necessarily whether enterprises should use AI.

It is:

As we allow AI to help us work, do we clearly understand what it can see?

And as for the lobster at the beginning of this article…

If there is data you really don't want it to see, protecting the data first is probably a good idea