Industry executives and experts share their predictions for 2026. Read them in this 18th annual VMblog.com series exclusive.
By Vipin Varma, Senior Vice President and Head of the Cybersecurity Practice, CitiusTech
A healthcare firm approached its technology partner with a rather unusual problem. The company had recently rolled out an Agentic AI system which used its SharePoint data to address employee queries. Now it faced a situation where all employees suddenly had access to restricted data, earlier available only to a closed user group. The company had to set stricter access control norms for the AI agent to ensure that it would not have unrestricted access to all company data.
Sixty-five percent of U.S. hospitals already use AI-assisted predictive tools as per the School of Public Health. According to Gartner, 29% of organizations experienced attacks on AI/GenAI infrastructure in past 12 months. Securing how these AI agents interact with data and with each other is as important as enabling regular human oversight when setting up an Agentic AI system. A malicious actor gaining access to the Agentic AI system could result in model poisoning or prompt injection attacks where user prompts are manipulated to trick the model into generating harmful outputs or deep fakes.
While Agentic AI systems can streamline healthcare delivery, there can be serious security ramifications if the required guardrails are not built in. Autonomous decision-making or unauthorized actions by the agentic system beyond its intended scope or adversarial attacks by malicious actors gaining access to the system are some of the biggest risks. As Agentic AI adoption becomes more widespread, it is critical for the organizational security policy framework to evolve to cater to these new complex risks.
Securing an Agentic AI framework is not a single step process and enterprises need to adopt a holistic approach, starting from the design and deployment stage, to monitoring and testing certain parameters on an ongoing basis once the system is deployed.
Critical safeguards must be built into the system at the development stage itself, like having a secure architecture for deployment of LLM-based systems and ensuring that it is compliant with industry defined security frameworks like OWASP, NIST and ENISA. This ensures that the system will be complaint with the required regulatory and industry norms while building in strong safeguards.
The first step to creating effective guardrails is defining the risk zones. After that, the guardrails and security control architecture should be designed based on best practices for that particular platform. Finally, these controls must be integrated and tested before they are operationalized.
Effective data governance enterprise wide is the next critical step. Identifying and classifying both structured and unstructured data, coupled with creating internal security policies on what data can be utilized and clearly defining what can be shared with specific user groups is essential. Building in control measures to provide user or role based authenticated access, with solutions for redaction or masking restricted data can help ensure that restricted data remains secure.
The third step would involve conducting adversarial testing to check for biases and hallucinations and that security measures are working as intended, and carrying out threat modelling must be done before the Agentic AI system is deployed. Practices like agent accountability logs and predefined containment protocols and autonomous agent drift detection can all help identify and control any security vulnerabilities in a timely manner. Parallelly, providing training to the teams that will be using and monitoring the Agentic AI system on how to use it effectively and red flags to look out for adds another critical layer of human oversight.
Finally, when it comes to Agentic AI deployments, there are two additional measures that come into play as compared to normal RAG based applications. The first is robust management of the non – human identities created along with governance of privileges and authorisations for access to resources. The other is monitoring and recording of all actions undertaken by the Agentic AI system for oversight and audit.
Once the system is deployed, continuously monitoring the output for any security compromise and using XAI or Explainable AI techniques to understand the output provided by the agentic system can catch issues like biases early on and enable corrective action in time. Measures like assigning AI agents unique identifiers and periodically recertifying their access and prompting human users for approval during sensitive or high-risk operations can all help reduce the risks associated with Agentic AI.
For healthcare firms, it is a question of when, and not if, they will deploy an Agentic AI system. Following best practices like data encryption and anonymization, rigorous model governance and security scans on an ongoing basis can go a long way in ensuring that your AI agent doesn’t go rogue.
##
ABOUT THE AUTHOR
Vipin Varma is the Senior Vice President and Head of the Cybersecurity Practice at CitiusTech. With over 28 years of experience in cybersecurity and broader ICT domains, Vipin brings deep expertise and strategic leadership to drive secure, resilient digital transformation across the healthcare ecosystem.
Prior to CitiusTech, Vipin led the cybersecurity business for global life sciences, healthcare, energy, and utilities clients at Tata Consultancy Services, where he spent over 12 years shaping cybersecurity strategies across regulated industries. Before entering the corporate sector, he served for more than 23 years in the Indian Army, where he held key roles including commanding a unit in active counter-insurgency operations, leading IT and communications for a UN peacekeeping mission in Sudan, and managing complex technology programs.
