The Hidden AI Security Risks Threatening Businesses

This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more. gettyAs businesses race to adopt artificial intelligence, new security risks are emerging alongside the technology’s benefits. These threats are challenging traditional security practices and introducing vulnerabilities that organizations may not be equipped to recognize or manage. As AI takes on a greater role in daily workflows and decision-making, leaders must understand how their organizations’ exposure is changing.

Protecting business systems and sensitive information requires more than adapting existing cybersecurity defenses; it also means considering the unique risks associated with AI technologies. Below, members of Forbes Technology Council highlight AI security risks business and tech leaders need to understand and share practical ways organizations can address them.

AI prompts can silently leak more sensitive information than most tech leaders imagine. For any AI tool, understand what goes into the underlying LLM, who hosts it and what privacy and security measures are provided. The more concentrated your LLM infrastructure or provider, the better your chance to guardrail your business’s sensitive information. – Alekh Jindal, Tursio

AI makes it very easy to find and exploit zero-day vulnerabilities in an operating system, hypervisor or anywhere in infrastructure. This means that the whack-a-mole game of patching systems as quickly as possible won’t work anymore. What tech leaders need to understand is that the game has shifted to prevention—the key is securing the infrastructure in a way that, even if there’s a zero-day, the data is still safe. – Ayal Yogev, Anjuna Security

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

The AI security risk leaders should watch is unchecked AI authority. Internally, as AI agents move from recommending to executing, organizations need clear permissions, policy-based authorization, approval thresholds and audit trails. Externally, AI accelerates vulnerability exploitation, making faster patching and continuous monitoring critical. The goal is not to restrict AI autonomy but to ensure it is governed at the point of action. – Shahir Daya, Zafin

Sending data, especially PII data and legal and compliance-sensitive data, through a direct API or MCP connection to a third-party model like OpenAI or Anthropic creates a new trust boundary. That data now flows through infrastructure you don’t control. Self-hosted models keep data and inference in one environment, reducing third-party exposure and the number of trust boundaries that can fail. – Julia Penfield, PH.D., VelocityEHS

Ask security how many AI tools and agents employees connected to company data last month—most can’t answer, and that blind spot is the real exposure, not the AI itself. Treat every AI tool like a company laptop: Give it an owner, log what it can access and revoke it the day someone leaves. – Siva Samy, ValGenesis, Inc.

One of the biggest risks is giving AI access to business data without fully understanding what it can access or how that data will be used. Companies need clear rules around data access, permissions and how AI-generated outputs are used. The goal is to build those safeguards into the technology from the beginning rather than trying to address them later. – Steve Cangiano, CMiC

AI agents act like humans, but they consume and move data far faster than any person could, often with more permissions than they need. Organizations need to govern them the way they govern insider risk: by understanding intent and behavior, not just access. In practice, that means connecting your AI security strategy to your data security strategy. Treat them as separate, and the gaps become the risk. – Sumit Dhawan, Proofpoint

Single-signal identity checks are obsolete. Fraudsters now use injection attacks and AI-generated media to bypass liveness checks, so a face or document that looks real proves nothing on its own. Verification must be continuous rather than a one-time onboarding checkbox, layering device and network signals, behavioral biometrics, and risk-based orchestration that pulls them together to build trust. – Kaarel Kotkas, Veriff

AI is turning hiring fraud into a real security threat. A fake candidate can now get through multiple stages of a hiring process, seem credible on paper and even ace interviews. If that person is hired, the company may hand an attacker legitimate access to internal systems, customer data, source code or financial information. Before giving candidates access to sensitive systems or data, companies need to verify that they really are who they say they are. – Gergo Vari, Lensa, Inc.

A widely underestimated risk is data leakage through shadow AI: employees quietly pasting sensitive customer, financial or IP data into unapproved tools whose retention, security and training practices IT cannot see. Bans can drive usage further underground. Organizations need better visibility to enforce responsible use and risk‑based controls that protect data without slowing innovation. – Neal Bradbury, Barracuda

While the focus is more on AI-based cost savings or productivity, emphasis on observability and compliance needs equal attention; if overlooked, it will lead to regulatory and credibility issues, along with an auditability nightmare. These vital layers decide the distance between a proof of concept and successful production of AI agents. – Shammy Narayanan, Welldoc

The biggest risk for me is not that AI will stop obeying humans but that humans will make a habit of trusting AI automatically. Although we are investing in better agentic AI defenses, we should also invest more in the internal architecture of the cyber professional, because the pressure will only increase, and without a zen-like mental clarity, the wrong choices could be made. – Carlo Tortora Brayda, Cyber Eagle Project Inc.

I don’t think enough leaders realize that AI security is fundamentally a data problem, not a model, agent or user one. Regardless of how AI is leveraged, it cannot operate safely unless it’s working with data that is trusted, accurate and high-quality. Teams can ensure their data meets these standards by implementing comprehensive, foundational data security posture management. – Thyaga Vasudevan, Skyhigh Security

Many product organizations underestimate the challenge of maintaining data access control as they scale AI. AI-augmented search, analytics and content synthesis offer enormous value, but preserving fine-grained access to sensitive IP is complicated if data is extracted into a data lake. Digital thread strategies need governed AI access to operational data and context to preserve access policy integrity. – Leon Lauritsen, Aras

One underestimated AI security risk is the rise of the “synthetic insider.” AI agents are given access to systems and permission to act for employees. They are no longer just tools but start to behave like insiders. Their actions should be continuously monitored against the role and task they were intended to perform. In the agentic era, leaders need to know whether an agent’s activity made sense. – Aviv Nahum, Above Security

Every doc uploaded to AI is a doc outside your retention policy. Company docs end up all over the place, including staff laptops. Staff upload docs because 1. there is no system with secure doc storage or 2. outside tooling is faster. Provide a fast and secure place for content that LLMs can access. Second, ensure docs are fully deleted. Ask AI vendors what happens to derived artifacts, embeddings, caches and backups when a source document is deleted. – Aaron Rallo, Trovia

One underestimated AI security risk is unclear accountability as AI moves from recommending actions to executing them. Organizations should define what AI can access, what it can act on autonomously, what requires human approval, and who owns the outcome. AI governance must address not only what a system can do but also who is accountable when it acts. – Jayasri Ranganathan, Trinity Solar Inc.

AI is collapsing the cost and time required for attackers to experiment. They can automate reconnaissance, generate thousands of attack variations, and continuously adjust based on what gets through. Static rules and known-pattern detection cannot adapt at the same rate. Organizations need behavioral and unsupervised detection to surface novel patterns, paired with real-time decisioning so newly identified risk can immediately inform controls. – Yinglian Xie, DataVisor

AI enables attackers to analyze stolen source code and cloud stacks at machine speed, exposing new vulnerabilities on the fly. Believing an attack is contained just by revoking initial access is untenable; block one vector, and attackers find another within hours. Businesses must move faster than feels comfortable, apply same-day patching, and treat self-inflicted downtime as the safer option. – Ryan Ikeler, MOXFIVE