A quiet office can look harmless. Racks of monitors bathed in light, headphones covering conversations, and the buzz of work carry on with no sign that something sinister lies underneath. But increasingly, there are accidental, unsanctioned technologies — a personal cloud folder here and an unsanctioned AI chatbot there. Soon, the organization will need to manage all of these new unanticipated risks. But shadow IT was just the first load of hidden threats. Shadow AI has upped the ante. A quiet office can look harmless. Racks of monitors bathed in light, headphones covering conversations, and the buzz of work carry on with no sign that something sinister lies underneath. But increasingly, there are accidental, unsanctioned technologies — a personal cloud folder here and an unsanctioned AI chatbot there. Soon, the organization will need to manage all of these new unanticipated risks. But shadow IT was just the first load of hidden threats. Shadow AI has upped the ante. What Shadow AI Is and Why It’s Growing What Shadow AI Is and Why It’s Growing An extension of shadow IT, shadow AI involves employees using non-approved technology. Shadow IT typically refers to consumer technology, like file-sharing apps or personal devices. Shadow AI usually involves fast-moving, data-hungry systems whose behavior can be erratic. Per research conducted by Gartner, 80% of organizations experience gaps in data governance. These gaps make it easier for people to miss AI-generated behavior. Many teams fail cybersecurity readiness assessments. The risk associated with AI is increased by employees adopting new tools faster than their teams can adequately review them. Since 30% of data breaches originate from vendors or suppliers, knowing what tools a team uses is a critical component of securing a company’s digital assets. 80% of organizations experience gaps 30% of data breaches originate from vendors Shadow AI has gained traction because employees view AI-generated content as a faster way to create content, summarize complex information, and troubleshoot technical issues. It reduces friction in daily work but introduces risks not previously seen with shadow IT concerns, including data exposure, compliance risk, and model-level risks. Shadow AI Versus Shadow IT Shadow AI Versus Shadow IT Shadow IT has long been blamed for unknown vulnerabilities. A high percentage of earlier breaches were due to unsigned SaaS tools or personal storage. AI tools change the equation entirely. The scale and speed at which they work, along with their opacity, create risks that are more difficult to detect and contain. With 78% of organizations utilizing AI in production, some breaches are now due to unmanaged technology exposure. The larger IT model still matters, but AI introduces a new dimension to broaden the attack surface. 78% of organizations utilizing AI Key Differences Between Shadow AI and Shadow IT Shadow AI is similar to shadow IT in that both stem from an employee's desire to be more productive, but they differ in where the risk resides. Shadow IT tools have fixed logic, which makes behavior predictable. Forecasting the behavior of shadow AI tools is more complex because models can be continuously modified and retrained.
Shadow IT risks include data being stored or moved without authorization. Shadow AI risks include model inversion, data poisoning, and model training.
Shadow IT is deterministic, while AI tools may hallucinate, generalize poorly, and overconfidently produce incorrect outputs. Shadow IT tools have fixed logic, which makes behavior predictable. Forecasting the behavior of shadow AI tools is more complex because models can be continuously modified and retrained. Shadow IT risks include data being stored or moved without authorization. Shadow AI risks include model inversion, data poisoning, and model training. Shadow IT is deterministic, while AI tools may hallucinate, generalize poorly, and overconfidently produce incorrect outputs. Shadow AI also arises in the context of upcoming regulations, such as the EU Artificial Intelligence Act, which could increase regulatory scrutiny. Security Risks That Make Shadow AI More Urgent Security Risks That Make Shadow AI More Urgent Shadow AI can lead to problems in engineering, marketing, and finance. As decisions are made based on AI outputs, proprietary data can be leaked, and internal business processes can be manipulated without anyone noticing. Model manipulation: Attackers can craft data that skews outcomes.
Prompt injection exposure: A created prompt can be used to extract private information from a model.
Data lineage gaps: AI tools may generate and store data in ways security teams can't track.
Compliance drift: AI tools change, and evolving governance plans may become irrelevant. Model manipulation: Attackers can craft data that skews outcomes. Model manipulation: Prompt injection exposure: A created prompt can be used to extract private information from a model. Prompt injection exposure: Data lineage gaps: AI tools may generate and store data in ways security teams can't track. Data lineage gaps: Compliance drift: AI tools change, and evolving governance plans may become irrelevant. Compliance drift: The concern grows with the advent of generative AI. A chatbot answering a vendor's question or a generative AI summary may seem harmless, but it risks revealing sensitive usage data or valuable proprietary intellectual property. Carnegie Mellon University found that large language models are far more vulnerable to adversarial prompts than rule-based systems. The problem increases when employees can use the tools without supervision. large language models are far more vulnerable An AI-enabled decision tree can be more biased than a conventional decision tree. Shadow AI often receives incomplete training information fed into third-party tools. Structured oversight of AI systems would ensure the integrity of updates. When teams overlook this, the model's data and behavior drift. How Security Teams Can Reduce Shadow AI Exposure Although shadow AI poses numerous risks, organizations can mitigate many of them by combining visibility with policy and technical controls, striking a balance that protects employee productivity without burdening them with time-consuming check-ins or blocked sites. Security teams benefit from treating shadow AI as a governance issue rather than a punishment issue. Mitigation strategies will inevitably need to evolve as employees use AI tools to improve productivity. 1. Build a Clear AI Governance Framework A governance plan should specify which AI tools to approve, what types of data employees can use, how to review model outputs before making high-stakes decisions, and what to do when an unpredictable model behavior occurs. The latter element includes who reviews the behavior, who investigates its causes, and what the consequences are. With oversight in place, organizations can treat AI as any other enterprise asset, subject to the same traceability, auditability, security, and compliance responsibilities as other legacy enterprise systems. 2. Provide Approved AI Tools Teams with access to vetted, centralized AI tools are less likely to turn to unapproved public AIs to bypass blockers. As jobs become more automated, staff will pour more effort into various models. Workers are already spending around 4.6 hours weekly using AI on the job, exceeding the average personal use time of 3.6 hours per week. AI from third parties, without proper monitoring, might already be more common than enterprise tools that are vetted and approved. Companies should take immediate steps to enforce their policies. 4.6 hours weekly using AI With a managed environment, organizations can monitor usage through tools, set permissions within databases, and enforce data governance across departments. This improves employee productivity while also protecting the business's data integrity and compliance. 3. Monitor Data Movement and Model Usage Visibility tools that flag abnormal behavior — such as sudden increases in AI usage, uploading data to unusual endpoints, or accessing the model in a short time frame with sensitive data — may help security teams identify misuse and data leaks. Reports indicate that over the past year, as many as 60% of employees utilized unapproved AI tools, and 93% admitted to inputting company data without authorization. 60% of employees utilized unapproved AI Detecting these patterns early may enable remediation, re-education, permission reconfiguration, or termination of the process before it leads to data leakage or compliance breaches. 4. Train Employees on AI-Specific Risks Cybersecurity training in general is not enough. AI can hallucinate by misinterpreting the intent behind prompts and generate seemingly authoritative, false, or biased content. Additionally, workers must understand that the use of AI differs from the use of software or services. Secure use requires changing mental models, understanding prompt risks, and handling personal data. hallucinate by misinterpreting the intent Users with basic machine literacy will fact-check output and be less likely to over-share personal data. They will treat the tools as valuable co-pilots, but they must be used under human supervision. Protecting Organizations Against Shadow AI Shadow AI is growing faster and is harder to identify than shadow IT. Although the scale and complexity of the risks differ, enlisting employee help can identify both more effectively. Governance policies can help companies strike the right balance. Security teams should reassess their exposure, stay vigilant for emerging threats, and act promptly before unseen AI-based tools make pivotal decisions in business applications.

The is an opinion piece based on the author’s POV and does not necessarily reflect the views of HackerNoon.

Should You Risk Using an AI Browser?

Portfolio

Nominated for 2022 - HackerNoon Contributor of the Year - Data Security

Is Shadow AI Worse Than Shadow IT?

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

10 Common Scams Targeting Healthcare Workers

The Noonification: Charting a Path for Global AI Regulation (11/14/2023)

AGI 2029: Walking the Fine Line Between Utopia and Existential Risks

Biden Charges Federal Government Agencies to Adopt AI

Biden Signs Executive Order to Support Workers Affected by AI

Biden Unveils White House AI Council to Drive Effective AI Policy Implementation

10 Common Scams Targeting Healthcare Workers

The Noonification: Charting a Path for Global AI Regulation (11/14/2023)

AGI 2029: Walking the Fine Line Between Utopia and Existential Risks

Biden Charges Federal Government Agencies to Adopt AI

Biden Signs Executive Order to Support Workers Affected by AI

Biden Unveils White House AI Council to Drive Effective AI Policy Implementation

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps