Shadow AI: The Silent Security Risk Lurking in Your Enterprise

Rachael ShahSenior Product Marketing Manager | F5

Artificial intelligence is here to stay. The gains in productivity, innovation, and economic impact are simply transforming how business runs across all teams. But the benefits come with serious risks for organizations adopting AI. One of the most urgent is shadow AI.

Shadow AI is the unsanctioned use of AI tools by employees without the knowledge or approval of IT and security teams. This practice is rapidly emerging as a significant security and compliance threat, even as it remains relatively new and continues to evolve.

Shadow AI is a fast-developing offshoot of shadow IT, the long-standing challenge where employees deploy unsanctioned software, cloud services, or systems within the enterprise. While shadow IT already strains governance and security, shadow AI adds a dangerous new layer of complexity. Unsanctioned software puts data in places IT can’t see; AI tools go further. Public models may ingest what employees feed them for training, making any exposure of proprietary information effectively irreversible.

A few examples of shadow AI:

• A software engineer uses a personal ChatGPT, Gemini, or Copilot account to generate boilerplate code, translate languages, refactor legacy code, or write test cases and unintentionally pastes proprietary code into a public model. That code could be stored, logged, or exposed. Worse, the AI-generated output might include insecure logic, fabricated APIs, or code that violates licensing terms.
• A communications specialist uploads a confidential strategy memo into an AI tool to summarize or draft a customer-facing message. That proprietary content is now sitting on third-party servers outside IT’s oversight.
• A sales rep installs a Chrome extension that auto-generates prospecting emails using AI and connects to both their email and CRM systems. This risks possible data leakage, internal policy violations, and noncompliance with regulations like GDPR or the California Consumer Privacy Act.

These examples are just the beginning. Countless other shadow AI scenarios are playing out every day across industries. The scale of the issue is so vast that Gartner® predicts that “by 2027, 75% of employees will acquire, modify or create technology outside IT’s visibility—up from 41% in 2022. As a result, top-down, highly centralized cybersecurity operating models will fail. Instead, CISOs must restructure cybersecurity into a lean, centralized function that supports a broad, federated set of experts and fusion teams embedded across the enterprise. This scales cybersecurity across the edge of the enterprise, closer to where technology and risk decisions are made and implemented.”¹

Prompt injection and data leakage: top concerns

While exact financial losses from shadow AI are still being calculated, the potential for harm is clear and growing. Organizations that fail to act now risk serious incidents, regulatory exposure, and erosion of customer trust.

Among the most concerning vulnerabilities is prompt injection—an attack where malicious instructions, supplied directly by an attacker or hidden inside content the AI processes, cause the model to bypass restrictions, leak sensitive data, or take unintended actions. AI models tend to trust input by default, making them vulnerable to such exploits. A successful prompt injection could trick AI into revealing internal data, corrupting automated processes, or undermining decision-making systems.

Another major worry is data leakage, especially of personally identifiable information (PII), protected health information (PHI), bank details and financial records, source code and proprietary IP, credentials and access keys, or customer records.

Compounding the problem is the potential for regulatory noncompliance. Unauthorized AI use can easily violate regulations and standards such as GDPR, HIPAA, PCI DSS, and the Sarbanes-Oxley Act, exposing companies to penalties and reputational harm.

Why encrypted channels hide unsanctioned AI

Part of what makes shadow AI so hard to control is that it travels the same way legitimate traffic does: encrypted. Nearly all web traffic now moves over TLS/SSL connections, and AI tools are no different. The chatbot in a browser tab and the API call behind an automation both ride encrypted sessions.

That encryption protects privacy, but it also creates a blind spot for traditional network security tools. A firewall that can’t inspect encrypted sessions sees only that an employee connected to an AI service. It can’t tell whether the employee asked for a lunch recommendation or pasted a customer database into a prompt. Legacy data loss prevention tools face the same limitation: sensitive information leaves the organization wrapped in sessions they can’t read.

The result is a gap in the organization’s security posture. Security teams can’t reliably detect shadow AI use, let alone stop it. Closing the gap requires visibility into encrypted traffic itself, with inspection at the points where prompts and data actually move, applied selectively so privacy and performance don’t suffer.

Navigating the regulatory horizon (EU AI Act and compliance costs)

The regulatory picture is tightening, too. The EU AI Act[he3.1] is phasing in: prohibitions have applied since early 2025, obligations for general-purpose AI models since August 2025, and further provisions, including transparency requirements, since August 2026, with obligations for high-risk AI systems following in 2027 and 2028. Penalties for the most serious violations reach up to €35 million or 7% of global annual revenue, whichever is higher. An unvetted tool that an employee wires into a business process falls under the same rules, whether or not anyone realized the law applied.

In the United States, the NIST AI Risk Management Framework is becoming the default playbook for demonstrating responsible AI governance. It’s voluntary, but auditors, insurers, and enterprise customers increasingly expect it. After an incident, regulators ask a simple question: did the organization know what AI was in use, and did it govern that use? Shadow AI makes the honest answer no.

The costs are measurable. IBM’s 2026 Cost of a Data Breach Report found shadow AI involved in 43% of security incidents, up from 20% a year earlier. Those breaches averaged $5.39 million and about one in five drew a regulatory fine.

Some companies respond with bans—but is that effective?

In response to these risks, some organizations have banned employee use of non-approved generative AI tools. Samsung, for example, was one of the first in 2023 to prohibit the use of ChatGPT and similar AI-powered chatbots, citing concerns over leaks of sensitive internal information. Several banks and other enterprises have since issued similar restrictions or warnings.

Yet industry experts overwhelmingly advise against blanket bans.

Why governance, not bans, is the better path

Despite the legitimate risks, industry analysts recommend preventive policies and governance frameworks over outright prohibitions. Most organizations haven’t gotten there yet. In IBM’s 2026 research[he5.1], 68% of breached organizations lacked AI governance to manage AI or detect shadow AI. Even among organizations with governance policies in place, only 29% regularly audited for unsanctioned AI.

Bans tend to be counterproductive. They’re difficult to enforce, they often suppress innovation and morale, and they can drive AI use further underground—making it even harder for security teams to detect and control.

As one IT leader said in a Gartner Peer Community discussion about banning AI tools: “Bans simply don’t work. Even without policies, this action hurts innovation and sends the wrong message to staff and the world about our organization.”

What organizations should do instead

Here are a few practical governance steps organizations can take right now:

1. Establish a clear AI usage policy
   Define what tools are approved, how data should be managed, which use cases are acceptable, and where guardrails exist. Make it crystal clear what’s permitted—and why.
2. Maintain a vetted list of AI tools
   Evaluate tools for data privacy, security, and regulatory compliance. Encourage employees to use these instead of unvetted platforms. For instance, use enterprise versions of tools such as ChatGPT with logging and data loss protections and adopt internal AI assistants integrated into company platforms.
3. Train employees and build awareness
   Help staff understand the risks—from data leakage and hallucinated content to compliance failures. Most misuse is unintentional.
4. Monitor, then adapt
   Consider using (ethically and legally compliant) monitoring to detect unauthorized use. Respond with education rather than punishment. Continually evolve your policies as AI capabilities change.

How F5 secures AI workloads and mitigates shadow AI

Can technology offer a scalable solution? Absolutely.

F5 AI Gateway sits between your users, agents, and the AI services they rely on. It delivers data leakage detection and prevention for every prompt and response that flows through it, blocking sensitive data (source code, credentials, customer information) before it reaches an external model. Paired with F5’s encrypted-traffic visibility solutions, which can steer AI-bound traffic to the gateway for inspection, that protection extends to the shadow AI tools hiding in encrypted channels (the default for most AI APIs and SaaS platforms).

AI Gateway also addresses the newer face of shadow AI: unvetted agents, plugins, and MCP servers operating inside approved tools. It governs which models and tools users and agents can reach at all, and it keeps a full audit trail of every interaction. Combined with the governance steps above, that turns an AI usage policy into something you can enforce.

Frequently asked questions

What is the difference between shadow IT and shadow AI?

Shadow IT is any technology employees use without IT approval, from unapproved apps to personal cloud drives. Shadow AI is its fastest-growing offshoot: the unsanctioned use of AI tools. The critical difference is what happens to data. Ordinary shadow IT puts data where IT can’t see it; AI tools can also ingest data for prompts and training, making exposure effectively irreversible.

How do IT teams detect unsanctioned AI in encrypted traffic?

IT teams detect unsanctioned AI in encrypted traffic by gaining visibility into that traffic itself. In practice, that means selectively decrypting and inspecting TLS sessions at network chokepoints and monitoring for connections to known AI services, with data loss prevention applied to prompts in transit. Without that visibility, security teams may see that an employee connected to an AI service, but not what data was shared, and AI features embedded in other SaaS apps may not be visible at all.

Why are blanket AI bans ineffective?

Blanket AI bans are often ineffective because a ban rarely stops the behavior. Employees who see real productivity gains keep using AI on personal devices and personal accounts, where security teams have no visibility at all. Governance built on approved tools keeps AI use where it can be seen and managed.

How does F5 AI Gateway prevent data leakage in prompt payloads?

F5 AI Gateway prevents data leakage in prompt payloads by inspecting prompts and responses in real time, spotting sensitive data in the payload (an access key, a customer record, a block of proprietary code), and blocking or redacting it before it reaches an external model. Policy controls also govern which AI models and tools users and agents can reach in the first place.

GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.

_{¹Gartner, Maverick Research: CISOs Must Transform Their Role or Become Obsolete, June 19, 2025}

About the Author

Rachael ShahSenior Product Marketing Manager | F5

More blogs by Rachael Shah