Skip to newsletter content
Will Lassalle
The Governance Gap

Issue #25

The "5% Power User" Threat Vector

For the past two years, the boardroom conversation around shadow AI focused on a simple, broad threat: regular employees pasting sensitive company documents into commercial LLMs like ChatGPT or Claude to write summaries.

ISSUE #25: The "5% Power User" Threat Vector

THE TAKE

For the past two years, the boardroom conversation around shadow AI focused on a simple, broad threat: regular employees pasting sensitive company documents into commercial LLMs like ChatGPT or Claude to write summaries. CISOs deployed web filters, drafted acceptable-use policies, and assumed the risk was managed. But fresh telemetric research published by Akamai in late August 2026 exposes a far more dangerous, concentrated reality: 5% of your workforce—your technical "power users"—are creating 90% of your catastrophic enterprise exposure.

The tension lies in the shift from casual use to hardcoded operational dependencies. While the average employee uses AI for quick drafting, these AI super-adopters interact with models at 12 times the rate of their peers. More importantly, they aren't just prompting chatbots; they are quietly integrating dozens of unvetted, specialized AI tools, browser extensions, and autonomous micro-agents directly into backend business operations. They are building custom scripts, linking APIs, and giving third-party models persistent access to core data streams—completely outside IT and security oversight.

As a CIO and CISO who has spent 25 years establishing enterprise risk frameworks, let me give you the direct bottom line: policing the broad workforce while ignoring your power users is a failure of security strategy. Your governance team is busy trying to control low-risk user prompts, while a hand-picked group of engineers and analysts are actively rewriting your corporate attack surface. When one of these long-tail, unvetted AI tools suffers a credential leak or supply-chain compromise, the blast radius will extend directly into your master cloud repositories.

THREE THINGS BOARDS SHOULD KNOW

  • Shadow AI risk is concentrated in technical super-adopters. — The top 5% of enterprise AI users create a disproportionate threat by embedding specialized, unmonitored micro-tools directly into core infrastructure.
  • Generic web filtering provides zero operational visibility. — Blocking public LLM domains does not stop power users from deploying localized, API-driven agents or unvetted developer extensions.
  • Power user workflows lack offboarding governance. — When an AI super-adopter leaves the organization, their hardcoded scripts and autonomous machine identities frequently remain active and unmonitored inside your network.

THE QUESTION I'M HEARING

"Our CISO reports that we have implemented enterprise licenses for major AI models and restricted unsanctioned web access. Why should the board be concerned about a small group of advanced users?"

You should be deeply concerned because advanced users do not limit themselves to approved corporate interfaces when those tools create operational friction. To bypass system limits, automate complex data pipelines, or speed up engineering cycles, power users routinely seek out specialized, long-tail AI tools that lack enterprise data protection agreements.

When an engineer or financial analyst links an unvetted third-party AI tool to an active database via an API key, they create an unmonitored, non-human identity with direct access to corporate intellectual property. If that specialized AI tool vendor suffers a data breach, or if the underlying model is manipulated via prompt injection, your corporate firewall will view the malicious exfiltration as normal, authorized system traffic executed by your employee's credentials.

My recommendation to the board is immediate and actionable: mandate a targeted "Power User AI Exposure Assessment." Direct your CIO and CISO to identify the top tier of high-volume data users across engineering, finance, and product teams. Instruct security operations to audit all active API tokens, custom browser extensions, and localized scripts assigned to these accounts. If your team cannot provide a live inventory of the third-party micro-tools operating within your environment today, your enterprise perimeter is fundamentally unmonitored.

THE BOTTOM LINE

Focusing AI governance on average employees while ignoring technical power users is defending the wrong perimeter. True enterprise resilience requires continuous visibility and strict access controls over the advanced micro-tools and hardcoded workflows quietly driving your business.