Shadow AI practices: A wakeup call for enterprises

This year marks a turning point in which we can say we are firmly in the era of AI agents, a revolutionary development in enterprise technology. Agents aren’t just a new software interface for enterprise processes, but a real technological advancement to boost the efficiency and scale of business operations.  

CIOs, CTOs or CISOs need to understand where agents sit on the maturity curve and how to leverage them to drive more profound transformation in their businesses. Companies that fully integrate agents into their workflows, as operators and teammates, will build the foundation for efficiency, quality and scalability to drive long-term growth and success.

AI agents draw on foundation models like OpenAI’s GPT, Google’s Gemini and Anthropic’s Claude to develop business reasoning and execution systems that learn and adapt.  It’s a serious leap forward when agents are combined with MCP (Model Context Protocol) servers, which connect agents to enterprise applications and data without requiring bespoke engineering and APIs.

Why AI agents benefit your organization

The core value of AI agents lies in handling complex, multi-step tasks with less human intervention.  They provide:

• Autonomy and efficiency. AI agents execute typical workflows, analyzing data, generating content and reports, sending alerts — faster than humans. This reduces manual effort by 50-80% in areas such as customer support and code development and debugging.
• Adaptability and reasoning. AI agents learn over time and with use. They link actions dynamically, learning from interactions to improve over time.
• Scalability for automation. AI agents are being deployed across industries like retail, e-commerce, media or business services, enabling 24/7 operations and handling high-volume tasks.
• Cost savings and ROI. AI agents speed up processes and deliver information faster and more reliably than humans in most cases. Studies show up to 40% productivity gains in knowledge work.

Where we stand on AI agent adoption

AI agents and their uses are slowly increasing across enterprises, but not as quickly as the hype suggests. A McKinsey report shows that many firms surveyed are still in the piloting or exploration stage. While 88% of firms are using AI in some form, only 23% are scaling agentic AI, and about 39% are experimenting with AI agents — primarily in IT, knowledge work or customer service.

Beyond these approved AI initiatives, employees may be going rogue, setting up their own internal LLMs and agents that bypass traditional IT policies and security measures. This should be a wake-up call for large enterprises: A tidal wave of risk may be about to hit you.

CISOs and CTOs need to start exploring ways to protect their organizations and establish new guardrails to stave off these threats. And even if you aren’t formally deploying AI agents, you need to be vigilant about shadow AI practices happening right now.

Industry insights from the field

In recent months, I’ve had numerous in-depth conversations with top CISOs and CIOs at industry events, including one group discussion where I spoke with a room of over fifty large enterprise executives. What topic was top-of-mind? The use of AI agents and AI security tools in their organizations.

What I learned from these executives was striking: There’s a significant gap between the AI market hype and organizational readiness. If you listen to the industry’s marketing, you’d believe AI agents are everywhere and every CISO is scrambling to buy solutions such as agentic security solutions, AI-specific firewalls or MCP lockdown products.  The hype has been further fueled by the fact that a handful of AI security start-ups have been acquired in the past six months by the likes of CrowdStrike, Palo Alto Networks, SentinelOne and CheckPoint.

Despite this prevailing belief, nearly all of the executives that I spoke with recently revealed that they had not yet deployed any of these innovative new solutions.  They were much more likely to have instead created processes and policies prohibiting or limiting AI usage, combined with newly implemented firewall rules on legacy systems. 

Most concerning, the threat of rogue AI usage, while troubling, was not being addressed as a burning near-term problem.  Unfortunately, from the conversations I’ve had with security service providers, rogue agents and MCP servers have sprung up in large numbers as employees try to test methods to perform their job with greater quality and ease.  These rogue agentic deployments could create a new set of security risks at multiple levels: To data, to traditional identity and access frameworks, to the AI agents themselves, or to AI hallucinations. Even worse, there’s the emerging threat of AI agents bypassing the boundaries set by human directives, to the detriment of the enterprise. 

Clearly, these risks should be at or near the top of the list of 2026 priorities for executives and boards.

Leveraging model context protocol

Model context protocol (MCP) is an open standard developed by Anthropic, and introduced in 2024. MCP servers are programs hosted locally or in the cloud that expose specific capabilities, tools, data sources or prompts to AI agents through a standardized, open protocol.

MCP is a secure communication standard allowing AI applications (as clients) to connect to these servers without custom integrations (or APIs) for each tool or data source. Servers provide three core building blocks:

• Tools. Executable functions and interfaces that AI models can invoke to perform actions, like querying a database or booking a flight.
• Resources. Data sources (e.g., files, APIs or real-time streams) that agents can access without computation.
• Prompts. Reusable templates to guide the AI’s behavior and allow for optimized, consistent use of the tools and resources.

While MCP — and industry alternatives that have been proposed — create a beneficial standard protocol for communications and interoperability, it does not address the security of the connection or access control privileges.  These functions must be handled by external solutions, similar to firewalls and IAM platforms in the TCP/IP world.  These solutions are starting to appear, with certainly more to come in the coming year.

One thing that worried me in my conversations: Executive awareness of MCP and its security implications is still limited. This needs to change in 2026.

Is agentic AI usage in line with the hype around AI investment?

As the McKinsey report shows, there’s still a lot of experimenting and tiptoeing going on with AI agents. From a CIO-CISO perspective, much of this is still informal exploration.

But the overall impression from news about AI investments suggests that enterprises are entirely on board with all AI offerings. And that simply isn’t the case. If you listen to a typical Silicon Valley startup, one might presume AI agent integration is exploding across the board. But it’s not.

To me, this may be a positive sign in some ways — it gives enterprise security teams time to catch up to the AI security reality. It also means there’s a lot of opportunity left for innovation, for vertical AI startups to meet the gap between the hype and actual use.

Clearly, there’s still substantial learning ahead for CISOs on AI agents and how to maintain secure guardrails for AI use, including rogue use, in the enterprise. Getting educated on AI agents and corporate security has to be at the top of the list of priorities for this year.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?