Sponsored by Akamai

As artificial intelligence rapidly transitions from an experimental novelty to a foundational engine of modern commerce, enterprises across every industry face a stark reality: embracing AI is no longer optional for maintaining a competitive edge. From large language models (LLMs) drafting code and analyzing market trends to autonomous AI agents executing multi-step business processes, intelligent automation is reshaping operations at a breathless pace.

However, this technological leap has outpaced the mechanisms designed to protect it. Traditional cybersecurity frameworks—historically built around static perimeters, predictable cloud traffic, and rigid data loss prevention (DLP) protocols—are struggling to keep up. As AI adoption accelerates, it introduces complex, multidimensional security threats that leave organizations uniquely vulnerable to data exposure, malicious manipulation, and severe regulatory compliance failures.

Recent research published by Akamai in its State of AI Usage Report 2026 and its corresponding State of the Internet (SOTI) Enterprise AI Usage Risk Report 2026 reveals a troubling disconnect between corporate oversight and real-world employee behavior. Nearly 20% of enterprise employees now interact with AI tools on a weekly basis, while more than 30% utilize them at least monthly. Crucially, a massive proportion of this usage occurs completely outside the visibility and oversight of corporate IT and security teams—a phenomenon rapidly evolving into the next major enterprise blind spot: shadow AI.

Main Facts: The Anatomy of Modern Enterprise AI Risk

The integration of artificial intelligence into daily workflows has fundamentally expanded the corporate attack surface. Unlike previous enterprise technologies, which served primarily as passive conduits or storage vessels for information, AI actively consumes, generates, stores, and evaluates enterprise data in real time.

According to Or Eshed, former CEO and co-founder of LayerX and current VP of Enterprise Security Products and Engineering at Akamai, the implications for enterprise risk are profound. "The blast radius is growing and growing, making it the path of least resistance for attackers nowadays," Eshed notes.

Key findings from Akamai’s 2026 research underscore the depth of the enterprise AI challenge:

• The Shadow AI Gap: Nearly half of all enterprise AI conversations occur via personal employee identities rather than corporate-managed accounts, obscuring data flows from IT departments.
• Vulnerable Browser Extensions: Over 16% of AI-related browser extensions analyzed in the SOTI report possess known Common Vulnerabilities and Exposures (CVEs).
• The "Power User" Dynamic: AI usage is heavily skewed; while the bottom 50% of enterprise users engage in 12 or fewer AI interactions, the top 5% generate at least 144 conversations, creating concentrated nodes of high-risk activity.
• Fundamental Architectural Shift: Traditional DLP solutions are built to monitor well-defined channels like email and cloud storage buckets, leaving them blind to the unstructured, conversational data movement inherent in modern LLMs.

Chronology: The Evolution of AI Threat Vectors (2025–2026)

To understand how rapidly the threat landscape has shifted, security professionals must examine the recent timeline of weaponized AI innovations and regulatory wake-up calls.

2025: The Rise of Client-Side Manipulation

As organizations rushed to integrate consumer-facing and specialized AI models, malicious actors quickly realized that standard browser interfaces offered a fertile ground for exploitation.

• Early 2025: Security researchers began documenting the vulnerability of AI-powered browsers and extensions. A landmark demonstration by LayerX highlighted how attackers could embed malicious prompt instructions directly into standard web pages.
• Mid-to-Late 2025: This technique gave rise to vulnerabilities such as "CometJacking," which targeted Perplexity’s AI-powered browser, Comet. By exploiting the naive nature of LLMs—which struggle to cleanly separate legitimate user intent from malicious page data—attackers successfully demonstrated how browsing an ordinary website could trigger unintended data exfiltration or unauthorized actions.

2026: Agentic Automation and Advanced Exfiltration

As the market shifted from simple chatbots to autonomous AI agents and developer-focused utilities, attack vectors evolved correspondingly.

• Early 2026: Akamai published its comprehensive State of the Internet Enterprise AI Usage Risk Report 2026, shedding light on the dangerous convergence of broad browser permissions and AI tools.
• Spring 2026: Security teams identified "CursorJacking," a novel attack vector weaponizing the extensive file and network permissions granted to the Cursor AI coding assistant. Threat actors silently harvested API keys, effectively impersonating legitimate developers and compromising entire software development lifecycles. This milestone proved that attackers no longer needed to crack perimeter firewalls; they simply needed to ride the coattails of trusted enterprise AI workflows.

Supporting Data: Unpacking the Numbers Behind Shadow AI

The numbers compiled in Akamai’s 2026 reports paint a vivid picture of decentralized, uncontrolled AI proliferation. While governance strategies across most enterprises remain strictly focused on reigning in a handful of well-known applications—such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google Gemini—employees are utilizing a sprawling ecosystem of hundreds of niche tools, plugins, and web-based applications.

Personal Identities vs. Corporate Accounts

One of the most startling discoveries of the Akamai SOTI study is that nearly 50% of corporate AI interactions take place using personal employee accounts. When workers use personal accounts for work-related research, drafting code, or analyzing financial reports, open models often ingest and train on that proprietary business data. Consequently, threat actors can leverage carefully crafted prompts in public AI tools to reconstruct sensitive corporate data, bypassing traditional network perimeters entirely.

The Threat of Prompt Injection

AI models are inherently "naive," possessing a limited capacity to distinguish between authorized instructions and malicious payloads hidden within text inputs. Prompt injection occurs when an attacker manipulates an LLM’s behavior away from the developer’s original design. When applied to AI agents—which are increasingly deployed across enterprise networks to execute decisions with minimal human oversight—the stakes multiply exponentially. Malicious actors can manipulate these agents via indirect prompts, enabling them to bypass authentication checks, scrape local directories, and exfiltrate connected sensitive files directly to external servers.

The Extension Epidemic

Browser extensions represent another blind spot for security teams. Employees frequently install AI assistants directly into their browsers to streamline productivity, unwittingly granting these extensions sweeping permissions to read web content, capture user credentials, and interact with private SaaS applications. With over 16% of these tools carrying known CVEs, they offer cybercriminals an easy backdoor into corporate systems.

Official Responses and Expert Insights: Redefining Security Strategy

Faced with this tidal wave of risk, industry leaders emphasize that traditional "prohibitionist" security strategies—such as outright banning AI tools or attempting to force an entire workforce onto a single approved platform—are doomed to fail. Such heavy-handed approaches inevitably drive employees further into the shadows, breeding resentment and hampering business productivity.

Or Eshed of Akamai stresses that organizations must adopt a multifaceted, risk-based approach to secure their AI estates.

"There’s AI you build in applications, AI consumed by the workforce, AI workloads as a cloud or compute unit. You need to address each and every one of them separately," Eshed explains.

Rather than trying to blanket-monitor every single interaction across the enterprise, security teams must direct their limited resources toward the specific users and tools that generate the highest exposure.

Data from the SOTI report illustrates this principle through the concept of AI "power users." Because the top 5% of users generate 144 or more AI conversations compared to the median employee, security architectures should prioritize monitoring these high-frequency environments. By deploying real-time warnings and policy guardrails at the exact point of interaction—such as when an employee attempts to paste sensitive customer data into an unverified public model—organizations can transform risky moments into valuable teachable opportunities without disrupting daily operations.

"We need to map all the applications, understand how sign-in works, optimize controls, recommend users to use the right AI application and turn day-to-day work into a teachable moment," Eshed adds.

Implications: The Path Forward for Secure AI Adoption

The unchecked explosion of enterprise AI usage carries profound implications for organizational security, data governance, and regulatory posture. If left unaddressed, the shadow AI visibility gap threatens to invalidate decades of progress in perimeter security and compliance management.

To successfully navigate this new era, security leadership must implement solutions that bridge the gap between innovation and defense. This is the precise challenge addressed by platforms like Akamai Workforce Protector.

Designed to operate seamlessly alongside traditional Security Service Edge (SSE), Cloud Access Security Broker (CASB), and Data Loss Prevention (DLP) solutions, Workforce Protector addresses critical visibility gaps at the endpoint and within the browser. By capturing interactions, prompts, full conversations, and agent runtime activities at the exact point of engagement, the platform empowers security teams to enforce granular policies in real time.

Whether an employee attempts to share sensitive source code with a public chatbot or an autonomous agent is improperly instructed to execute an unauthorized financial transaction, modern workspace security tools can guide, warn, or block the action instantaneously. Crucially, this is accomplished with zero-disruption deployment, ensuring that productivity remains high while corporate assets stay protected.

Ultimately, the future of enterprise security belongs to organizations that can harmonize visibility with adaptability. By acknowledging the realities of shadow AI, understanding the mechanics of prompt injection and browser extension vulnerabilities, and deploying targeted, context-aware security platforms like Akamai Workforce Protector, enterprises can confidently scale their AI initiatives while safeguarding their most critical data assets.