_{This content is sponsored by Akamai}

AI tools, from large language models (LLMs) to AI agents, are becoming a fundamental component of business operations, and organizations across all sectors recognize that they must embrace AI to stay competitive.

However, the rapid uptake of AI has introduced new security challenges that many organizations are struggling to address, and traditional security approaches are unable to keep pace with these risks.

Akamai found in its State of AI Usage Report 2026 that nearly 20% of employees use AI tools on a weekly basis, while more than 30% use them at least monthly. These include tools that sit outside of the visibility and oversight of IT teams – a phenomenon known as shadow AI.

Another recent report, Akamai’s State of the Internet (SOTI) Enterprise AI Usage Risk Report 2026 highlighted the unique nature of AI and the risks it introduces: “Unlike previous technologies, AI continuously consumes, generates, stores, and acts upon enterprise data. The downside is that AI introduces entirely new categories of risk that traditional security tools were never designed to address.”

The adoption of AI across enterprise environments is creating new attack vectors and expanding the attack surface.

“The blast radius is growing and growing, making it the path of least resistance for attackers nowadays,” said Or Eshed, former CEO and co-founder of LayerX and current VP of Enterprise Security Products and Engineering at Akamai.

As AI adoption accelerates, security teams require new capabilities to monitor AI activity, enforce policies and mitigate emerging threats. The ability to gain visibility and control across the environment is essential for secure AI adoption.

At the same time, these measures must avoid creating unnecessary friction for employees or disrupting business operations.

Understanding the AI Threat

The risks posed by enterprise AI usage are manifold, stemming from both accidental insider threats and malicious threat actors.

Here are some of the key AI-related risks security teams must understand to confidently adopt and scale AI technologies while preventing data exposure.

Shadow AI Risks

Shadow AI presents a substantial data security risk to enterprises. This is in part due to the sheer volume of AI applications that are publicly available, beyond well-known platforms such as OpenAI’s ChatGPT, Anthropic’s Claude and Google Gemini.

The Akamai SOTI report noted: “For many organizations, AI governance remains heavily focused on managing these few well-known platforms – creating a significant visibility gap with regard to the true scale of decentralized AI adoption across the company”

Eshed explained that employees often use their personal accounts across various AI apps for both work and non-work purposes.

What’s more, Akamai’s study found that nearly half of all enterprise AI conversations take place through personal identities rather than corporate-managed accounts.

This is problematic as these open models will train on business data and understand the employees’ personally. This mix can create opportunities for threat actors to obtain sensitive data if they use the right prompts in public AI tools.

Prompt Injection

While AI tools have become incredibly powerful, capable of executing tasks far quicker than humans, they are also highly open to manipulation.

Eshed described AI models as “naïve”, less able to distinguish between legitimate and malicious requests compared to humans.

Prompt injection occurs when an input into an AI tool alters the model's behavior in ways the application developer did not intend. The technique is typically used by malicious actors to steal sensitive data – either from the models themselves or by using them as a gateway to access other parts of the target’s network.

In 2025, LayerX demonstrated how an attacker could manipulate a browser’s AI by embedding malicious instructions within a web page, leading to the potential exposure of sensitive data and perform actions the user never intended. The vulnerability was dubbed CometJacking as it affected Perplexity's AI-powered browser, Comet.

A particularly effective target for prompt injection attacks are AI agents, which are being rapidly deployed across network architecture and can execute decisions with minimal to no human oversight.

The Akamai SOTI report highlighted how threat actors can manipulate agents with prompts to bypass traditional user authentication entirely and exfiltrate sensitive local data or connected files directly.

The Growth of AI Browser Extensions

As AI adoption rises, there has also been a significant growth in employees accessing AI through browser extensions.

These extensions often have broad access to web content, user activity, credentials and enterprise applications.

AI browser extensions, particularly those with elevated permissions, represent huge opportunities for attackers to access sensitive data or take over user sessions.

This can be achieved through vulnerability exploitation – the SOTI study found that 16.31% of AI extensions have known CVEs.

In addition, threat actors are developing novel techniques to exploit AI extensions. An example of this is an attack vector known as “CursorJacking”, identified by Akamai researchers in 2026.

Threat actors weaponized broad permissions often granted to everyday extensions to silently exfiltrate API keys from the Cursor AI coding assistant tool. This allowed them to impersonate developers and compromise their entire AI-driven development environment.

The Data Governance Challenge

Even across embedded, approved LLMs, data governance is a huge challenge for security teams. Employees increasingly share unstructured data with AI models in a highly fragmented way, including prompts, conversational context, source code and screenshots. Collectively, these interactions could lead to the models obtaining sensitive business data they are not meant to hold.

The SOTI report noted that traditional data loss prevention (DLP) controls are not set up to manage such data movement, instead designed for well-defined channels such as email, file transfers and cloud storage.

The Solution: A Risk-Based AI Security Strategy

Given the scale of AI deployment, it is simply not possible, or desirable, for defenders to block these tools or force all employees to use a certain model.

Instead, a targeted, risk-based approach, is the most effective way to manage AI risks without impacting business productivity.

Eshed said that organizations must accept there is no one single security tool that can address AI risks throughout an enterprise.

“There’s AI you build in applications, AI consumed by the workforce, AI workloads as a cloud or compute unit. You need to address each and every one of them separately,” he explained.

Security teams should also focus on the AI tools and users that pose the biggest risk to the business.

“In reality you have a small subset of users and activities that create a lot of risk,” Eshed noted.

For example, the SOTI report identified the phenomenon of “power users” within AI usage. The research found that while the average enterprise user participates in 36 AI conversations, the bottom 50% of users engage in 12 conversations or fewer. In contrast, the top 5% of users generate at least 144 conversations.

Security controls should be focused on these power users by monitoring their prompts and sending real-time warnings when they are engaging in potentially risky behaviors – like inputting sensitive customer data − at the point of interaction with models.

In some cases, individual employees may need to be blocked from using AI for work purposes.

AI governance resources should be directed toward users who present the highest levels of AI-related risk and exposure.

Visibility around the use of AI across the enterprise is key to developing a targeted approach to AI governance.

Eshed said: “We need to map all the applications, understand how sign in works, optimize controls, recommend users to use the right AI application and turn day-to-day work into a teachable moment.”

Why Organizations Should Consider Akamai Workforce Protector

Akamai’s Workforce Protector offers security teams a foundation upon which to build a risk-based approach to AI security.

Workforce Protector tracks AI usage across network architecture and works alongside traditional Security Service Edge (SSE), Cloud Access Security Broker (CASB), and data loss prevention (DLP) tools to identify and address critical gaps in AI tools.

It is a workspace security platform that addresses browsers and AI applications. Workforce Protector provides real-time visibility into employee engagement with AI, alongside SaaS, web, and private applications, at the exact point of interaction.

“Inside the browser, we’re able to catch interactions, prompts, full conversations, connected across any application. We’re able to inspect the interaction of users with AI applications on the device, as well as the runtime activity of the agents,” Eshed explained.

This analysis is used to determine the risk of different AI tools and users’ engagement with them, allowing security teams to select appropriate guardrails and enforce policies at the point of interaction.

Such policies include guiding, warning and even blocking users in real time when potentially risky actions are attempted.

Eshed noted: “We can force them to obey corporate guardrails. If you ask an agent to go and do something on your behalf, such as wire money somewhere, we will be able to prevent that.”

The solution offers zero-disruption deployment with controls only utilized when necessary and ensuring minimal impact on productivity.

Because Workforce Protector is embedded in the lifecycle of AI applications on devices and browsers, it does not impact to an organization’s day-to-day operations.

Conclusion

The use of AI in enterprises has exploded beyond security teams’ ability to control the significant data security risks these tools present. The scale of shadow AI today means many of these tools are not visible to organizations, making data governance challenging.

Security teams must gain oversight over AI usage across their networks and apply controls at the point of interaction to balance strong security with productivity.

Akamai’s Workforce Protector can provide the foundation to such a strategy, offering visibility into all AI tools used in the enterprise, and enabling teams to enforce appropriate controls to prevent damaging data security issues occurring.