Unmasking shadow AI

Recent studies from IBM and IO have highlighted the increased use of shadow AI in organisations and the associated risks to data privacy and security.

 

Security often gets sidelined in a rush to adopt new technologies, especially when their use is not properly managed. Artificial Intelligence (AI) is no exception: IO’s research revealed that the unsanctioned use of AI tools is now a growing problem for businesses in the UK and the US. More than a third (37%) of the surveyed organisations admitted to employees using AI without permission or guidance.

 

The IBM study, meanwhile, showed that 63% of the surveyed organisations did not have AI governance policies to prevent the proliferation of shadow AI or manage the use of AI. The findings also illustrated the risk, calculating that having a high level of shadow AI added an extra 511,000 GBP to the global average cost of a breach.

 

Both studies make an important point: Shadow AI has become pervasive, exposing businesses to potential data leaks, compliance breaches and reputational damage – and a lack of oversight is to blame.

 

High level of usage, little oversight

In many organisations, individual employees spearheaded AI adoption long before their employer was ready to officially embrace the technology. Without their employer’s knowledge and explicit consent, workers are harnessing free, easily accessible tools such as ChatGPT and Otter.ai to research information, create or summarise notes and complete writing tasks. With Google searches now showing results generated by its AI, Gemini, some employees might even be using AI without realising it.

 

Understandably, users love the convenience and efficiency that AI tools provide. But, shadow AI has created unknown business risks, as some users employ third-party tools in ways that their organisation does not consider acceptable.

 

Where generative AI (GenAI) is used for business purposes, data exposure, privacy and security are the main concerns. Whether it’s for generating text, creating extracts or building a presentation, employees often share sensitive corporate information with the AI. This could include financial data, confidential documents or even full email trails.

 

If this information falls into the wrong hands, it can lead to significant damage. Just imagine a company roadmap being exfiltrated, exposing strategic information to competitors.

 

Equally, cyber-criminals who gain access to company emails could gather insights into user behaviours, writing styles, communication patterns and more. This would enable them to generate highly convincing phishing emails – impersonating business contacts or responding to real conversations – to launch sophisticated ransomware and other cyber-attacks.

 

Add to that the fact that AI isn’t always right. The quality of its output depends on the quality of the data input, so an AI tool might return inaccurate or outdated results, leading to bad business decisions.

 

Agentic AI raises the stakes

Whenever a technology is adopted outside of company-managed environments, there is a lack of visibility, verification and oversight. An organisation that doesn’t know what AI tools are being accessed or what data is shared with them has no way of implementing controls and governance to minimise the risks.

 

While it could roll out guidelines on what users should and shouldn’t do, the nature of shadow AI means that it cannot enforce them.

 

Agentic AI – systems that can make autonomous decisions and act on them – will only exacerbate this problem.

 

While the core concern currently is employees unwittingly revealing sensitive company data, bigger threats are to come. Once agentic AI becomes embedded into work processes and is allowed to operate independently, it could cause more severe damage.

 

To carry out tasks without human interaction, agentic AI needs direct access to data, enterprise systems and other applications through protocols like MCP (Model Context Protocol). Through these interfaces, AI that is allowed to act autonomously could make unwanted changes to a company’s systems. For example, an AI tool that connects with a database to facilitate advanced business intelligence queries could extract and forward large amounts of data far beyond what the user intended to share.

 

However, while agentic AI is expected to become ubiquitous in business environments, it will be somewhat easier to control. These applications will typically need to be configured or managed by an IT team, giving companies better visibility into their use.

 

Don’t ban AI – manage it

AI is not going away. As these applications become more common, they will cause greater security concerns.

 

But, there are ways to mitigate the risks associated with AI, from rolling out authorised applications and monitoring their use to ensuring human oversight.

 

In the early days of Bring Your Own Device (BYOD) and cloud applications, organisations learnt an important lesson: If you don’t give users the tools and features they want, they will go and find them elsewhere.

 

This is why, just like with BYOD, banning AI outright won’t work. Instead, businesses must take the opposite approach – giving employees a happy path with secure access to approved tools, as well as appropriate guidelines and user training. 

• Any AI tool should be vetted properly before roll-out to ensure it meets the company’s security standards, and to understand its constraints. Evaluate the risks and determine what risk levels you are willing to accept.
• Understand the regulatory requirements around AI. Consider the flow of data and make sure it complies with GDPR and other requirements.
• Deploy approved tools under corporate management and monitor how they are used.
• Put safeguards in place and ensure you have the security basics covered – such as strong authentication processes. In the IBM survey, 97% of breached organisations that had experienced an AI-related security incident admitted they did not have proper AI access controls in place.
• Publish clear internal guidelines for the use of approved AI applications – and if necessary, take steps to prevent access to unsanctioned tools. Because AI is evolving fast, regularly review and update the guidelines.
• Training must be mandatory. Educate staff on the safe, responsible and effective use of AI. Remind them of the risks AI can pose and teach them not to input sensitive data without following the correct guidelines. 

Shadow AI is a silent threat. The risks range from cyber-security and compliance concerns to competitive exposure. Luckily, organisations that implement authorised tools, guidelines and training can still harness AI effectively while managing these risks.

 

Proactive governance, rather than banning AI, is the answer. When there is an official path, employees are far more likely to follow it.

 

---

 

Nadir Merchant is General Manager, IT Operations Suite at Kaseya

 

Main image courtesy of iStockPhoto.com and pixdeluxe