Top 5% of AI Users Create 12x the Security Risk, Akamai Finds

Enterprise AI security teams are watching the wrong threat

While most organizations focus on preventing employees from using ChatGPT for routine tasks, a small group of AI power users is quietly embedding unvetted tools into critical business operations—and creating outsized security risks in the process.

According to research from Akamai, the top 5% of enterprise AI users interact with models at 12 times the rate of the bottom 50% of the workforce. These super-adopters routinely engage in conversations lasting 18 prompts or more, compared to the five-prompt average for typical employees. The difference signals that AI has moved beyond productivity assistance to become an embedded collaborator in essential business functions.

"Small groups of AI power users are casting outsized shadows across enterprise threat surfaces that are already riddled with dips and blind spots," said Or Eshed, Vice President of Enterprise Security Product & Engineering at Akamai. "While security teams are focused on trying to govern all employees' access to big frontier LLMs, the cumulative long-tail shadow of dozens of smaller AI tools used by power users arguably poses a more significant security risk."

The shadow AI problem

The visibility gap extends far beyond frontier models. Nearly half of all enterprise AI conversations—47.11%—occur through personal identities rather than corporate-managed accounts, according to Akamai's State of the Internet: Enterprise AI Usage Risk Report 2026.

The divide between managed and unmanaged AI is stark. Platforms with dedicated governance controls like Gemini Enterprise (98.15%) and Microsoft Copilot M365 (90.55%) keep most interactions inside corporate identity systems. Meanwhile, DeepSeek (99.8%), Microsoft Copilot Standard (63.92%), ChatGPT (61.36%), and Claude (61.09%) are dominated by personal logins.

Akamai found that 14.4% of enterprise AI conversations occurred via corporate email addresses linked to personal freemium AI subscriptions. This means sensitive data entered into prompts may be used for public model training, even when employees use company credentials.

Browser and IDE extensions represent another rapidly expanding blind spot. At midsize enterprises, 17.7% of employees use at least one AI extension, compared with 9.53% at larger organizations. Nearly 75% of these extensions request high or critical permissions, and 16.31% contain known CVE vulnerabilities—significantly higher than the 10.80% vulnerability rate for browser extensions overall.

New attack vectors emerge

This expanding AI surface is creating novel attack methods. Akamai's report highlights vibe hacking, where attackers modify local instruction files to manipulate AI coding assistants into generating vulnerable code. CursorJacking involves weaponizing rogue extensions to harvest API keys and session tokens from local databases. CometJacking uses indirect prompt injection embedded in malicious web pages to trick AI agents into exfiltrating local files.

Why it matters

The concentration of AI usage among power users means security risk is not evenly distributed across the enterprise. Organizations that treat AI governance as a universal employee problem may miss the concentrated threat posed by the small percentage of workers who have made AI tools central to their workflows. Security teams need to identify which employees depend most heavily on AI, understand which tools they're using outside approved channels, and determine whether those systems operate within enterprise guardrails—before adversaries exploit the gaps.

The findings were first reported by The Hacker News, based on Akamai's research and telemetry data from real-world enterprise usage.