As organizations look to embrace a new generation of autonomous AI agents which can access enterprise systems, make decisions and interact with other AI agents, new questions are emerging around how these systems are governed and secured.
As the governance of agentic AI and non-human identities rapidly one of the most significant challenges organizations will facein the coming years, we spoke to Umashankar Lakshmipathy, EVP and Head of Cloud, Infrastructure, and Security Services, EMEA, Infosys, to find out more.
- As organisations move from AI assistants to autonomous AI agents, what new security and governance challenges are emerging?
The move towards autonomous AI agents creates new security and governance challenges because these systems can make decisions and take actions with limited human involvement. Unlike AI assistants that primarily generate information or recommendations, agents can access enterprise data, use applications and tools, communicate with other agents and act on behalf of people or business processes.
Latest Videos FromTechRadar
This greater autonomy increases the potential business impact of an error, manipulation, or security compromise. An agent may have a valid identity and legitimate system access but could still take action that falls outside its intended purpose. The risk becomes more significant when agents operate across several systems, as one incorrect decision can trigger a sequence of actions that are difficult to detect, explain or reverse.
Organisations must therefore govern more than the AI model. Every agent needs a defined business purpose, an accountable owner and clear limits on the data, systems, and actions available to it. Sensitive or high-impact activities require stronger safeguards, while organisations need sufficient visibility to trace an outcome back to the person or process that initiated it, the agent that acted and the tools it used.
This represents an important evolution in AI governance. Organisations must continue to manage the quality and safety of AI-generated content, but they must now also control the decisions and actions that autonomous agents can execute. Getting that balance right will be essential to achieving the productivity benefits of agentic AI without creating unacceptable operational, financial, or regulatory risks.
- As agentic AI adoption grows, why are “non-human identities” becoming such an important issue for enterprises? What are the risks and opportunities when AI agents begin communicating and making decisions with other AI agents?
Non-human identities are becoming an important enterprise issue because a growing proportion of digital activity is performed by applications, APIs, service accounts, automation and AI agents rather than by employees. AI agents increase the significance of these identities because they can interpret information, make decisions and initiate actions, rather than simply execute predefined instructions.
This makes visibility, ownership, and access management essential. Organisations need an accurate inventory of the non-human identities operating across their environment, together with a clear understanding of their business purpose, accountable owner and access to systems and data. Without this foundation, excessive permissions, long-lived credentials, and identities that remain active after their original purpose has ended create material security exposure.
Communication between agents creates significant business opportunities. Specialised agents can work together to automate complex processes across customer service, operations, software engineering and cyber defense. A process that previously required several teams and systems could potentially be coordinated by a group of agents operating at much greater speed.
However, agent-to-agent collaboration also creates chains of dependency and trust. A compromised or poorly governed agent could pass inaccurate information through a workflow, influence another agent’s decisions or attempt to use a more privileged agent to perform an action that it could not execute directly. These risks can multiply when organisations cannot establish which agent initiated an action or under whose authority it was performed.
Organisations should therefore enable agent collaboration within defined business and security boundaries. The identity, authority and purpose of each agent should be evaluated independently rather than allowing trust or access rights to pass automatically from one agent to another. This approach allows enterprises to benefit from multi-agent automation while preserving control and accountability across the complete business process.
- How can organisations strike the right balance between giving AI agents autonomy and maintaining appropriate oversight and control?
The appropriate level of autonomy for an AI agent should be determined by the business impact of its activities, the sensitivity of the information involved and how easily its actions can be reversed. Applying the same level of oversight to every agent would either create unnecessary friction or leave higher-risk activities insufficiently controlled. For example, an agent that analyses public information or produces an internal summary can operate with relatively light oversight. An agent that accesses customer information, changes production infrastructure, makes regulatory decisions or initiates financial transactions requires much tighter boundaries.
However, organisations should introduce autonomy progressively. Some agents will remain assistive and provide recommendations for people to consider. Others may prepare actions that require human approval. More mature agents can operate independently within approved limits, escalating situations that fall outside their normal authority. Agents performing highly predictable and well-controlled activities may eventually be given broader autonomy, provided their actions remain visible, governed and capable of being stopped.
Human oversight remains important, but requiring approval for every action would undermine the efficiency that agentic AI is intended to deliver. It can also lead to approval fatigue, where reviews become routine and provide little meaningful protection. Human involvement is most valuable for actions that are financially significant, legally sensitive, difficult to reverse or outside the agent’s established operating boundaries.
The aim is not to give agents the maximum possible autonomy. It is to give each agent enough autonomy to deliver business value while keeping its decisions within an acceptable level of risk. Routine and reversible activities can be automated, while consequential decisions remain subject to stronger safeguards and clear organisational accountability.
- Are existing cybersecurity, identity and governance frameworks sufficient for the age of agentic AI, or do organisations need a new approach?
Existing cybersecurity, identity and governance frameworks provide a strong foundation for agentic AI, but they are not sufficient without further adaptation. Organisations do not need to replace established capabilities such as Zero Trust, identity governance, privileged access management and AI risk management. They do, however, need to apply these capabilities more dynamically to agents that can make decisions and take actions at machine speed.
Many traditional controls were designed around granting access to a person, application or service and reviewing that access periodically. Agentic AI introduces a more dynamic operating model. An agent may access several data sources, invoke multiple tools, interact with other agents and make numerous decisions during a single business process. Authentication alone cannot establish whether every subsequent action is appropriate.
Existing frameworks therefore need to be extended in several practical ways. Every agent should have a unique identity, an approved business purpose, an accountable owner and a managed lifecycle. Its authority should be limited to the task it is performing and should not remain active for longer than necessary. Stronger safeguards should apply when an agent accesses sensitive information, uses privileged tools or attempts an action with a significant business impact.
Organisations must also be able to reconstruct how an outcome was produced. This means maintaining a clear record linking the originating person or business process to the agents involved, the data and tools they used, the actions they took and the resulting outcome. This visibility is essential for oversight, investigation, regulatory assurance and rapid intervention when something goes wrong.
The required approach is therefore an evolution of existing security and governance rather than the creation of a separate control environment for AI agents. By extending established frameworks to govern agent identity, authority, and actions continuously, organisations can scale agentic AI while maintaining coherent enterprise-wide control.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
