Addressing the AI Security Chasm
Addressing the AI Security Chasm
How to secure AI starts with knowing where it is being used, understanding the risks it creates, applying the right controls, keeping those controls current, and validating that they work. As AI adoption accelerates, doing all five is becoming increasingly important.
AI is already embedded in business applications, internal workflows, third-party products, and autonomous systems. But adoption is moving faster than many security programs can inventory new AI capabilities, evaluate their risk, or establish the controls needed to protect them.
The result is an AI security chasm: a growing gap between using AI, securing it effectively, and being able to demonstrate that it is secure.
Closing that gap requires more than AI governance or a list of best practices. Organizations need visibility into where AI is being used, a clear understanding of the risks it creates, controls designed to address those risks, and credible evidence that those controls are working.
The Growing Gap between AI Adoption and Security
AI introduces a visibility challenge before it becomes a control challenge.
Organizations may develop their own AI systems, enable AI functionality within existing technology, allow employees to use generative AI tools, or rely on vendors that incorporate AI into their products. Each creates another place where organizational data, systems, and processes may interact with AI.
Then there is shadow AI, or AI tools and features being used without the security team's approval or knowledge. When security teams cannot see where AI is operating, they cannot consistently evaluate the data it can access, the actions it can perform, or the risks it introduces.
As HITRUST has previously noted, “AI adoption will continue to accelerate. The question is whether assurance can evolve at the same pace.”
But visibility is only part of the challenge.
Governance can establish policies, responsibilities, and acceptable use. It does not, on its own, demonstrate that the controls protecting an AI system are operating effectively. In other words, organizations need to address both sides of the chasm: controlling AI risk and proving those controls work.
Moving from AI Visibility to AI Security
The first step in securing AI is understanding what needs to be secured.
Organizations should maintain visibility into internally developed models, generative AI tools used by employees, AI agents, AI features introduced into existing software, and third-party AI that interacts with organizational data.
Once those systems are identified, security teams can evaluate the AI security risk associated with each one.
That means looking beyond the underlying technology. Organizations should consider model and third-party dependencies, data exposure, access controls and permissions, integrations with business-critical systems, and attacks designed specifically to target AI.
HITRUST's 2026 threat findings have highlighted risks including user execution, AI-enabled phishing and deepfakes, and “Exfiltration via AI Agent Tool Invocation.”
Agentic AI makes this especially important. If an agent can interact with internal systems, security teams need to understand what it can access, which actions it is permitted to take, and whether manipulated inputs could cause unintended behavior or expose sensitive information.
Not every AI use case carries the same level of risk. Organizations need to assess AI systems based on factors such as data sensitivity, system access, autonomy, and business impact, so that higher-risk systems receive greater scrutiny.
For organizations that need a defined approach for evaluating AI governance and risk management practices, the HITRUST AI Risk Management Assessment provides a way to assess those practices.
Turning AI Security Best Practices into Testable Controls
Understanding risk is important. But understanding a risk does not reduce it on its own.
Organizations need AI security controls that translate identified risks into specific safeguards. That is an important distinction when considering AI security best practices. A best practice can describe what an organization should aim to achieve. A prescriptive control defines what needs to be in place in a way that allows its implementation and effectiveness to be tested.
Depending on the AI system and its risks, those controls can address access and permissions, data protection, model and supply chain security, input and output filtering, monitoring, and human oversight.
LLM security best practices should address what information a model can access and how inputs and outputs are handled. Controls such as input and output filtering can help reduce the risk of malicious prompts, unsafe responses, or inappropriate disclosure of sensitive information.
Generative AI security best practices should also consider how prompts, uploaded data, and generated outputs are handled. Depending on the use case, organizations may require human review before AI-generated content reaches customers or influences important decisions.
Agentic AI security best practices need to account for the ability of agents to take action. Strong permission controls, human-in-the-loop oversight, segmentation of AI components, and input and output filtering can help limit what an agent can access and do.
An AI security framework can provide a structure for putting these protections in place. But not every framework serves the same purpose. Most focus primarily on governance or risk management, while few provide prescriptive controls that can be tested and certified.
The question is not simply whether an organization follows an AI security framework. It is whether the approach provides the controls and evidence necessary for the risks the organization faces.
AI Security Has to Evolve with the Threat
Even well-designed controls cannot remain static while AI and the threat environment continue to rapidly evolve and change.
A control can continue operating exactly as designed while becoming less effective against new or evolving threats. As HITRUST has previously stated, “A control can still be operating exactly as designed and still no longer be sufficient for the risk it was intended to address.”
This is why assessing AI system security is not a one-time exercise.
Security requirements need to evolve alongside the threat landscape. Through its Cyber Threat Adaptive approach, HITRUST uses threat intelligence to evaluate whether its requirements continue to address the techniques being used by threat actors. According to HITRUST's Q2 2026 analysis, threat coverage was over 97% for the AI Security assessment.
Organizations should not only ask whether controls continue to operate. They should also ask whether those controls remain appropriate in response to an evolving technology and threat landscape.
Independent Validation Closes the AI Security Chasm
Putting controls in place is essential, but it does not completely close the AI security chasm.
An organization can say it follows AI security best practices. It can establish policies, implement controls, and conduct its own testing. But that is different from having those controls tested by someone other than the organization that implemented them.
Independent validation provides evidence beyond an organization's own assertions.
The HITRUST AI Security Certification brings together prescriptive, AI-specific controls, independent validation, and centralized quality assurance, scoring, and certification. It can be completed as a standalone assessment or added to an e1, i1, or r2 assessment.
That distinction matters for organizations trying to secure AI systems while also demonstrating security to customers, boards, regulators, and other stakeholders.
As the HITRUST Trust Report states, “With a 99.62% breach-free rate among HITRUST-certified environments, HITRUST stands alone in cybersecurity assurance.”
Closing the AI security chasm ultimately requires organizations to move beyond, saying the right controls are in place. They need credible evidence that those controls have been tested and are operating effectively.
The AI Security Chasm Extends to Third Parties
Organizations also need to consider the security of the AI systems they are buying.
Third-party applications and services increasingly incorporate AI capabilities, which means an organization's AI exposure can extend beyond the systems it develops or operates directly.
The same questions apply. Where is AI being used? What data or systems can it access? What risks does it introduce? What controls are in place? And what evidence demonstrates those controls are working?
Nearly half of breaches now reach the enterprise through the vendor and supplier ecosystem: the 2026 Verizon Data Breach Investigations Report found that 48% of breaches involved a third party, a 60% year-over-year increase. Source: Verizon 2026 DBIR.
As AI becomes more prevalent across third-party products, relying solely on vendor questionnaires or self-attestation may leave organizations without the evidence they need to make informed risk decisions.
Independently assessed evidence can provide greater confidence that a vendor's AI security controls have been implemented and validated.
Closing the AI Security Chasm
AI adoption is not slowing down, and organizations cannot afford to wait for the technology to stabilize before addressing the risks it creates.
Understanding how to secure AI systems requires visibility into the AI use across an organization, a clear understanding of the risks those systems introduce, prescriptive controls that address those risks, and a way to keep those controls aligned with an evolving threat landscape.
But control alone is not enough.
Closing the AI security chasm also requires credible evidence that those controls are operating as intended. Independent validation helps organizations move beyond simply stating that their AI is secure toward demonstrating that the controls protecting it have been tested.
Ready to strengthen your organization's approach to AI security? Get started with HITRUST.
