Companies are not doing enough to protect their data. Instead, they are mainly reacting after a problem occurs, but not anticipating or being proactive before an issue starts. That’s the word from a number of industry experts who are sounding the alarm on AI applications and agents running on unstructured data that threatens to swamp even the most well-designed enterprise security infrastructures.
“Most are doing security theater on the technical side and nothing at all on the human side,” said Jon Wolheim, founder of Wolheim Group. “They buy tools, write a policy nobody reads, and call it governance. Right now most data isn’t being stolen out the front door. It’s being walked out the side door by well-meaning employees.”
The challenge? According to Mona Rajhans, software engineering manager at Palo Alto Networks, “AI adoption has expanded the data attack surface faster than governance has kept up. Policies written 2 years ago did not anticipate agentic workflows, model APIs, or employees pasting sensitive data into consumer AI tools.”
Most companies have implemented baseline security in terms of encryption, controlled access, and systematic audits.
However, “Few companies have advanced security mechanisms that integrate AI ethics, lifecycle management, and oversight of cross-border overflow,” said Leela Lakshmi Jayaram, hardware-in-the-loop engineer at Ford Motor Co. “Even with those, the main challenge in today’s digital and AI-focused world is the fragmented accountability of data security measures among different units—IT, legal, and business. Companies work with many third parties whose data processors and AI model training principles are not visible.”
If anything, the gap between security spend and actual protection is widening, said Dan DeCloss, founder and chief customer brand officer at PlexTrac. “Companies have built up fancy security stacks, but many still can’t tell you which vulnerabilities and exposures are genuinely putting their sensitive data at risk, versus which ones are noise on a scanner report. Most companies still haven’t solved that problem, but they have gotten better at generating more findings to be confused about.”
In addition, many companies do only what is minimally necessary to meet regulatory and other legal requirements, agreed Rebecca Herold, CEO and founder of The Privacy Professor and an IEEE member. They need to “treat data security and privacy as fundamental components of comprehensive enterprise risk management,” she said. “Compliance alone does not necessarily mean that an organization understands where its data resides, who has access to it, how it is being used, or what would happen if that data were compromised.”
The gaps are pronounced across two areas: unstructured data and AI. “[T]raditional security and governance practices are straining to keep pace with the rampant growth of unstructured data,” according to the authors of a survey from the Cloud Security Alliance. The survey “found that while three-quarters of organizations are confident in their ability to secure unstructured data, more than two-thirds (68%) report that less than 80% of their unstructured data is protected. …” In a separate survey, only 27% of cybersecurity leaders said they could track more than 75% of their sensitive data. This Bedrock Data survey of 530 organizations also found that 70% identified AI and machine learning governance as top security concerns.
The New Data Threatscape
“Everyone recognizes that today’s AI landscape delivers enormous benefits— but it also brings substantial risks,” said Jayaram. “For individual users, those risks are concerning; for companies, they are exponentially greater. Organizations hold highly sensitive business information, intellectual property, plus strategic data that can be exposed or exploited with surprisingly little effort by competitors or hackers.”
Such threats “are less exotic than people expect,” said Rajhans. Leading threats include “sensitive data flowing into external model training pipelines, agentic systems querying internal databases without a human in the loop, and shadow AI tools ingesting proprietary information nobody approved.”
One of AI’s greatest security risks is the way it can dramatically expand a company’s attack surface. “When employees put company data into these tools, you’re potentially giving them to external parties,” said Aimee Simpson, director of product marketing at Huntress. “Especially if you’re using an AI tool without sufficient guardrails, you could leak data without realizing the extent of your problem.”
Only about 25% of companies attempt to remediate critical vulnerabilities, Herold said. Added to this is a growing threat from AI. “One in four malicious breaches were AI-enabled. Organizations are confronting increasingly sophisticated threats while fundamental weaknesses remain.”
This threatscape is also amplified by what has become an “explosion of nonhuman identities created by the rapid development and adoption of agentic AI,” said Rakesh Shah, VP of product management at Quest Software. “For every human identity, there are now dozens of service accounts, APIs, and autonomous agents acting independently. Identity and access security have become central to AI success. If governed data is the guardrail for what AI knows, identity security is the guardrail for what AI does.”
Shadow AI is another rising challenge to cybersecurity. “People paste customer lists, contracts, and source code into consumer chatbots because it makes their afternoon easier, and no firewall sees any of it,” said Wolheim. “Just look at the millions of artifacts people inadvertently made public when learning how to create artifacts with Claude; they’re all fully indexed and searchable on Google now. Imagine if those millions of breaches were each compromised customer financials, HIPAA-protected heath data details—even government secrets.”