# How to protect corporate data when using AI agents

*At the TechCrunch Disrupt conference[a discussion took place](<https://youtu.be/4FZTk2J4tc0?si=66-NsRldhCKVwDCy>)focused on the security of AI agents. Participants discussed the risks of deploying autonomous AI systems, access management to corporate data, infrastructure protection, and new requirements for developing such solutions. Here are the key takeaways.  

***Zak Whittaker, Security Editor at TechCrunch**

**Gavriel Cohen, Co-founder and CEO at NanoCo**

**Rick Smith, President of Products and Technology at Okta**

**How to ensure the security of AI agents**

AI agents are capable of performing tasks that previously required human involvement. However, connecting to various tools, applications, and corporate data creates additional risks.

One way to mitigate these risks is to run agents in an isolated environment and provide access only to the resources necessary for a specific task. Critical actions, such as sending documents to external recipients, may require human confirmation.

At the same time, companies must apply the same access management principles to AI agents as they do to employees. Without clear restrictions, agents may gain access to applications and data that are not needed for their work.

**What mistakes companies make**

When implementing AI, organizations often focus on increasing productivity, addressing security issues only after the systems are launched. Companies do not always predefine which data, applications, and tools an agent should have access to.

To reduce risks, the principle of least privilege is used—an agent receives only the permissions necessary to perform a specific task.

An additional layer of protection is verifying actions against user intent. Before starting work, the agent creates a plan that is approved by a human. Then, a separate model checks whether the execution of the task aligns with the agreed-upon plan.

Restrictions within the application alone are not enough. If an agent escapes the isolated environment, it should not automatically gain access to the rest of the infrastructure.

**Why old vulnerabilities are becoming more dangerous**

Autonomous AI agents can explore infrastructure and systematically exploit system weaknesses. Therefore, unpatched vulnerabilities may pose an even greater threat to companies.

To protect against this, it is necessary to regularly update software, fix known vulnerabilities, and verify which resources are accessible to agents when they exit a container or virtual machine.

It is particularly difficult for organizations with limited IT budgets to ensure such protection, as they may not always have time to update infrastructure and implement additional security measures.

**Why banning AI does not always help**

A complete ban on connecting AI to corporate data can lead to the opposite result. If employees cannot use agents to work with internal systems, they may start manually transferring data to external services.

During the discussion, an example was given of a large company whose employees were unable to connect agents to the corporate repository. To create analytical dashboards, they independently exported tens of thousands of customer records and uploaded them to the web version of Claude for processing.

This approach increases the risk of information leakage. A safer alternative is controlled access to necessary data with the ability to monitor the actions of AI agents.

Companies must also consider the risk of intentional misuse of AI by employees. To address this, it is important to control data transmission, network connections, and interaction with critical systems.

**Why agents need an emergency stop button**

In the event of a threat, a company must be able to quickly stop the operation of an agent. This requires mechanisms for revoking permissions, disconnecting individual tools, and terminating access to corporate resources.

If an agent escapes the isolated environment, revoking data access alone may not be sufficient. In such cases, it will be necessary to stop computational processes and restrict access to the infrastructure.

However, many companies still run agents on regular work computers without full isolation. This complicates monitoring their actions and emergency shutdown.

**Teams of AI agents**

Companies are gradually shifting from using individual AI assistants to systems where multiple agents collaborate to perform tasks and exchange information.

In the coming months, teams of AI agents may become one of the key areas of industry development. The number of agents working within a single team could increase from dozens to hundreds, and major research labs are already experimenting with systems that include thousands of agents.

At the same time, each agent must have a responsible owner who controls its operation. As these systems evolve, companies will need to ensure secure interaction not only between agents and employees but also between agents themselves and external commercial services.

**What developers need to consider**

The security of AI agents requires regular software updates, fixing known vulnerabilities, and controlling access to corporate resources.

When developing products, it is also important to consider the pace of AI advancement. While a corporate solution goes through approvals and security checks, technologies and customer needs may change significantly.

Therefore, developers must design systems taking into account not only current tasks but also requirements that may arise in the coming months. This will make products more flexible and prepare them for new scenarios of AI agent usage.
