Generative AI adoption in U.S. businesses is accelerating, even as security vulnerabilities multiply at an alarming rate. That leaves a stark gap between productivity gains and the infrastructure meant to protect them. A Wharton School and GBK Collective study tracking 1,416 enterprise leaders and employees found that 63 percent of workers using AI said it helped them finish tasks faster. At the same time, researchers uncovered nearly 268,000 internet-facing systems running AI-related products and exposed to potential breaches—a 63 percent increase in nine months, from October 2025 to July.
On the surface, the productivity story looks promising. Forty-six percent of employees said AI helped them produce higher-quality work, and 47 percent said it freed them to spend more time on work they found interesting. Among frequent AI users, 52 percent held jobs that met job-quality standards, compared with 32 percent of employees who never used AI. But those advantages aren't evenly distributed. Forty percent of college graduates used AI at least weekly, compared with just 17 percent of employees without college degrees. Managers used the technology far more regularly than individual contributors.
The Infrastructure Problem
That’s where the optimism breaks down. Censys researchers found companies deploying AI infrastructure at breakneck speed without adequate security controls. The most exposed product was n8n, a workflow automation platform that connects AI agents to corporate applications and services. Censys counted nearly 114,000 internet-facing n8n deployments, including roughly 77,000 running software versions affected by at least one recently disclosed critical vulnerability. Researchers also identified more than 54,000 exposed control panels for OpenClaw, an open-source AI agent.
“It's a little bit beyond just like the tool itself, but these [tools] talk to some really sensitive systems,” said Emily Austin, a principal researcher at Censys. Censys CEO Zakir Durumeric pointed to the scale of the problem: companies already struggled to track existing software before employees began building their own AI applications. These shadow deployments can connect to emails, financial records and internal databases—the very crown jewels adversaries target.
The risks aren't theoretical. Authorities arrested and charged Saudi national Abdulrahman Alfozan, 35, after he allegedly installed an audio recording device in an OpenAI office in San Francisco. Alfozan had worked at OpenAI since July 2025. Around September 30, someone found a concealed audio recording device inserted into a USB port on the back of a television in an executive conference room. Federal authorities said security footage showed Alfozan near the executive conference room multiple times, though he had no authorization to enter the space or attend meetings.
Federal prosecutors said a chatbot account linked to Alfozan's spouse, which shared key identifiers with Alfozan's account, submitted prompts between September 24 and September 30 seeking information on assembling and deploying surveillance hardware. Those prompts asked for technical details on building continuous audio recorders, customizing Wi-Fi connectivity, improving live listening, encrypting SD cards and concealing remote-access devices. The company later determined someone had installed the device two days before its discovery and that its contents were encrypted.
FBI Director Kash Patel wrote on X that the bureau and its partners arrested Alfozan after he allegedly concealed a recording device inside his employer's executive conference room “to secretly capture sensitive communications about the company's super intelligence tech.” Patel added: “Protecting American SI innovation is critically important and FBI views it as one of our highest counterintelligence priorities.”
Threats Beyond Today
Perhaps more unsettling, Anthropic's Frontier Red Team warned that smaller AI models may be only months away from spreading independently across internet-connected household devices, including smart ovens, TVs and other electronics. Researchers demonstrated an AI model running across 15 inexpensive Android TV boxes. Separate experiments, they said, showed more advanced models replicating across simulated networks. Unlike traditional malware, self-replicating AI could adapt to unfamiliar devices and find new ways to spread without human intervention. Researchers cautioned that they hadn't yet demonstrated a fully autonomous AI worm operating in the real world. Still, they warned continued advances could make such threats increasingly difficult to contain.
The Responsibility Question
Employers bear responsibility for these gaps. Among employees already using AI, 31 percent said their use had led employers to ask them to take on more responsibilities—potentially without greater authority or compensation. Fifty-two percent of all employees said they had less influence over technology adoption than they wanted. Christos A. Makridis, an associate research professor at Arizona State University and senior researcher at Gallup, recommended that employers provide paid time to learn and test approved tools on actual job tasks. He also urged employers to explain permitted uses, clarify how they will check outputs and encourage deliberate experimentation. Training should include frontline workers, not just AI pilot volunteers. Employees should be able to flag errors or question proposed uses without being labeled resistant to innovation.
OpenAI said it discovered the recording device and used its own models to identify the employee responsible. The company said it found no evidence that Alfozan accessed or exfiltrated model weights, source code or customer data. OpenAI terminated him. Austin recommended companies treat AI infrastructure like any other deployed software and review, update and secure it regularly.
Why This Matters:
The AI deployment surge reveals a critical governance failure in American enterprise. Companies are racing to capture productivity gains without basic security hygiene, leaving vulnerabilities that hostile state actors and industrial spies are actively exploiting. The Alfozan case isn't an anomaly; it's a warning. When infrastructure expands faster than oversight, adversaries exploit the gap. AI benefits also remain uneven across education levels and job types, suggesting market forces alone won't solve the productivity puzzle. Some companies and workers will capture gains while others fall further behind. For policymakers and business leaders, the takeaway is clear: faster deployment without security discipline doesn't create competitive advantage—it creates liability. The window to establish proper governance frameworks and training protocols is closing rapidly.