Big security problems rarely announce themselves. Often, the real danger comes from small pockets of behavior that slip through existing controls. In the case of enterprise AI, recent research points to an uncomfortable pattern: a tiny fraction of users can create an outsized AI security risk—not through malicious intent, but through how frequently they use AI and the kinds of tools they bring into daily workflows.
Akamai’s State of the Internet: Enterprise AI Usage Risk Report 2026, based on real-world telemetry and threat analysis, suggests that the top 5% of AI “power users” interact with AI models at a rate far higher than most employees. That difference reshapes the risk profile of an organization, turning “occasional” shadow usage into a persistent stream of data exposure and new attack opportunities.
The concentration problem: 5% driving 12x activity
Security teams often focus on governing employees who use popular AI tools for quick drafting or brainstorming. While that effort is important, the research highlights a different driver of risk: heavy users who embed AI more deeply into business operations.
According to Akamai data, the top 5% of enterprise power users interact with AI models 12 times more frequently than the bottom half of the workforce. The pattern shows up in conversation length as well. Where the average employee may run about five prompts per interaction, power users routinely engage with 18 prompts or more. In practice, that means more iterations, more context being shared, and more moments where sensitive information can be included.
Why that matters for security teams
When AI use concentrates among a small group, it becomes easier for those workflows to “hide in plain sight.” Instead of scattered, low-volume usage, the organization faces repeated access patterns tied to the same individuals and potentially the same toolchain. Akamai’s experts describe this as an “outsized shadow” across enterprise threat surfaces that already contain gaps and blind spots.
Shadow AI expands the attack surface
Even organizations that manage well-known enterprise AI platforms may struggle to see what happens outside approved channels. The report points to a growing ecosystem of additional tools—niche AI apps, personal subscriptions, and AI-enabled SaaS—that employees adopt without IT oversight.
This is where the AI security risk grows quickly. Shadow AI increases opportunities for data leakage and can introduce autonomous AI agents operating in the enterprise but outside established guardrails.
Akamai compares this dynamic to device behavior: employees increasingly “bring their own AI tools” (BYOAI). The analogy matters because the result is the same as unmanaged devices—visibility gaps around how data is stored, retained, and processed.
Personal logins create visibility gaps
One of the most striking findings is not just which AI tools people use, but how they access them. Nearly half of all enterprise AI conversations—47.11%—occur through personal identities rather than corporate-managed accounts.
That creates a sharp contrast between AI platforms that enforce corporate identity boundaries and those that rely on personal accounts. For security, the difference is simple: corporate governance typically gives IT, security, and compliance teams stronger visibility and control.
Examples of governance vs. personal dominance
Some platforms show clear governance dominance, with the vast majority of interactions staying inside corporate identity systems. Others are heavily dominated by personal identity logins.
- Gemini Enterprise (98.15%) and Microsoft Copilot M365 (90.55%) largely keep interactions within corporate-managed identity.
- DeepSeek (99.8%), Microsoft Copilot Standard (63.92%), ChatGPT (61.36%), and Claude (61.09%) show a strong tilt toward personal identity logins.
The takeaway for an organization is that “AI adoption” is not a single category. It is a combination of model choice, account type, and identity management practices—each affecting where data goes and who can monitor it.
Corporate email used with personal subscriptions
Governance can also become muddled when employees register personal AI subscriptions using corporate email addresses. In such cases, logs may look “corporate,” but the license and training pathway may behave more like a consumer service.
Akamai reports that 14.4% of enterprise AI conversations occurred via corporate email addresses linked to personal “freemium” AI subscriptions instead of enterprise-managed licenses. That matters because sensitive data injected into prompts may end up used for public model training, depending on the service’s policies.
For security leaders, this is a reminder that identity signals aren’t enough. You also need clarity on licensing context and how data handling works for the exact subscription type employees use.
Long-tail blindness: extensions and niche tools
Traditional governance often targets a shortlist of mainstream AI products. But the report suggests employees are quietly adopting dozens of smaller AI tools, AI-enabled SaaS applications, and personal subscriptions that expand the ecosystem beyond IT visibility.
Browser and IDE extensions are a particularly fast-moving blind spot. At midsize enterprises, 17.7% of employees use at least one AI extension, compared with 9.53% at larger organizations. Importantly, nearly 75% of these extension requests seek high or critical permissions.
Vulnerable extensions increase exposure
Permissions are one concern, but known software vulnerabilities are another. Akamai found that 16.31% of AI extensions contain known CVE vulnerabilities, versus 10.80% across browser extensions overall. That difference suggests AI-related add-ons may carry a higher operational risk profile than general extensions.
From a threat perspective, these tools can create unmanaged pathways into active user sessions and sensitive corporate data. Akamai frames the consequence broadly: this Shadow AI landscape isn’t only a privacy issue—it can become the infrastructure that enables automated cyberattacks.
Where attacks bypass traditional controls
As the AI surface grows, it also creates new ways for adversaries to operate outside traditional protective patterns. The report outlines several emerging attack categories that leverage how AI tooling behaves inside real environments.
Vibe hacking
Attackers can subtly modify local instruction files (for example, AI configuration documents) to influence AI coding assistants. The goal is often to push the assistant into generating vulnerable code or performing actions it should not.
Cursorjacking
Rogue extensions can be weaponized to silently harvest sensitive material such as API keys, session tokens, and proprietary source code from local databases.
Cometjacking
Adversaries can use indirect prompt injection embedded in malicious web pages. In these scenarios, the “target” shifts from the human endpoint to the AI collaborator, tricking AI agents into exfiltrating local files.
A CISO checklist to reduce AI security risk
If the mission is to reduce the AI security risk, the answer is not only stricter policies for a few headline tools. Akamai’s guidance focuses on building control points that match how AI is actually used—by teams, extensions, and even agents.
1) Establish continuous visibility
Discover AI applications, browser/IDE extensions, and agents across the network. Then inspect what matters in real time—prompts, uploads, and responses—so you can see exposure as it happens, not after the fact.
2) Eliminate shadow AI through identity enforcement
Enforce corporate SSO, block unmanaged personal logins, and audit corporate email addresses tied to “freemium” subscriptions. This approach closes the visibility gaps created by personal identity usage.
3) Use contextual AI DLP at the prompt level
Deploy prompt-level inspection to catch sensitive leakage in unstructured content. Legacy pattern-matching tools can miss the nuance of code snippets, internal text, or other context-rich prompt content.
4) Audit extensions and permissions rigorously
Maintain an inventory of AI-related browser and IDE extensions. Enforce strict permission boundaries and screen add-ons for known CVEs so risky tools don’t become a default part of employee workflows.
5) Govern AI agents like privileged identities
Treat autonomous AI agents and AI-enabled browsers as privileged digital identities. Apply least-privilege access, strict scope limitations, and continuous monitoring so agent behavior stays within acceptable guardrails.
Conclusion: act on concentration, not just adoption
Enterprise AI adoption is accelerating, but risk doesn’t scale linearly with headcount. Akamai’s research suggests that a small group of power users can create a disproportionate AI security risk by using AI more intensely and by relying on tools, accounts, extensions, and agent behaviors that sit outside conventional governance.
To protect the organization, security teams need to shift from “are employees using AI tools?” to “where exactly is AI operating, which workflows depend on it, and are those systems constrained?” With continuous visibility and stronger controls over identities, extensions, and agent permissions, enterprises can reduce the outsized shadow before it becomes an entry point for attackers.
Source: https://thehackernews.com/2026/08/the-outsized-shadow-why-5-of-ai-users.html
