AI is changing the cybersecurity operating environment because increasingly autonomous systems can take actions on behalf of people and organizations. They can write code, make recommendations, trigger workflows and interact with data and applications at a speed traditional security processes were never designed to match.

That shift framed a recent CyberLive conversation on LinkedIn featuring Rishi Mehta, Chief Information Security Officer at HCLTech and Amit Jain, Executive Vice President and Global Head of Cybersecurity at HCLTech. Their discussion focused on a central question for the AI era: how can organizations build enough trust in increasingly autonomous systems to use them at scale without losing control of risk? 

AI expands the cyber operating environment

Frontier AI gives attackers greater speed, scale and accessibility, but the speakers cautioned against focusing on model capability alone when assessing the threat. Skilled operators can combine existing models with the right surrounding tools and workflows to produce increasingly capable outcomes, making the broader system around the model an important part of the equation.

That combination of models, tools and human expertise can support everything from phishing and social engineering to script development and activity inside a compromised environment. The speakers expect attackers to make greater use of AI capabilities already deployed within targeted organizations, potentially using those systems to locate sensitive information, support lateral movement and accelerate post-exploitation activity. At the same time, many of the same capabilities can strengthen defense, with security teams applying AI across vulnerability management, incident response, governance, compliance, red teaming, penetration testing and product security. 

This creates an environment in which both sides can use AI as a force multiplier. Jain highlighted how that shift also expands the attack surface itself, particularly as Agentic AI introduces a growing population of machine identities capable of interacting with systems, data and business processes. As organizations deploy more agents, identity and access management becomes increasingly important as the population of non-human identities inside the enterprise grows.

Identity becomes a critical control point

Many existing identity programs were designed around people, devices and applications rather than large populations of autonomous agents. This becomes problematic when an AI agent inherits the identity and permissions of the person who created it. Human users are often granted broader access than they require for every individual task, so transferring those same permissions to an autonomous agent can give it unnecessary reach across data, applications and business systems.

Organizations will increasingly need to assign machine identities according to the purpose and authority of each agent, with clear visibility into who owns it, what information it can access and which actions it can perform. That becomes even more important as shadow AI grows, because agents and AI-enabled capabilities can enter the organization faster than traditional security inventories and approval processes were designed to accommodate.

This makes visibility and control fundamental. Organizations need to understand what AI is operating in their environment and whether the risks associated with those systems remain within defined boundaries. This extends beyond chatbots to agents, low-code and no-code capabilities and AI embedded in third-party software, all of which can create additional pathways into enterprise systems. 

Governance moves into runtime

Identity is one part of a broader governance challenge. Organizations often put significant effort into decisions made at the beginning of an AI initiative, including the model, data, hosting environment and initial controls. Jain argued that the same level of attention needs to continue after deployment because AI systems operate continuously, interact with changing environments and can behave differently as conditions evolve.

“The governance and guardrails need to scale into runtime,” said Jain. 

Runtime governance means continuously observing behavior, enforcing policies and maintaining the ability to intervene when a system moves outside expected boundaries. Depending on the use case, that could mean limiting permissions, requiring human approval, rolling back an action or stopping an agent altogether. An initial inventory and posture assessment remain important, but they provide only a snapshot. Trust in production needs to be supported by evidence gathered as the system operates.

“Evidence-based trust is what I would clearly recommend instead of implicit trust,” said Jain. 

That evidence also needs to reflect the level of authority granted to the AI system. An internal agent performing a low-risk administrative task does not carry the same consequences as one operating in a customer-facing process, supply chain or core financial system. Testing, approval thresholds, monitoring and rollback capabilities should become more rigorous as authority increases, while consequential actions require clear traceability so organizations can understand which agent acted, what information informed the decision and who ultimately owns the business outcome.

“Eventually, it's the organization that owns the outcome, not the algorithm,” said Jain. 

The security workforce is changing with the technology

The same AI capabilities introducing new risk can also help security teams manage an operating environment that is becoming faster and more complex. 

The speakers pointed to the evolution of the security operations center analyst as an example. Analysts once had to gather investigative evidence manually from different systems before they could understand an incident and respond. AI can now perform much of that initial analytical work, allowing the analyst to begin with greater context and focus attention on investigation, judgment and response.

Malware analysis has followed a similar path. Automation reduced much of the basic reverse-engineering workload while specialist roles remained important for more complex problems. AI is extending that progression by accelerating analysis further, allowing skilled professionals to spend more time on work requiring creativity, experience and contextual understanding.

The broader direction is toward security operations where AI handles more repeatable analysis and humans supervise, investigate complex cases and make higher-order decisions. This matters in an industry that has long faced skills constraints, but Mehta stressed that cybersecurity professionals will also need stronger AI literacy themselves. Understanding how the technology behaves, experimenting with it and recognizing its limitations are becoming important both for securing AI systems and for using AI effectively as part of defense. 

Resilience becomes part of AI security design

As AI becomes more deeply embedded in business processes, organizations also need to assume that systems will occasionally behave in unexpected ways. Models can be manipulated, prompts can be exploited and agents can make decisions that produce unintended consequences. That reality is pushing security leaders to place greater emphasis on fail-safe design, recovery and resilience alongside prevention.

Jain argued that familiar cybersecurity principles remain highly relevant. Secure by design, defense in depth, Zero Trust and continuous assurance still provide the foundation, but they now need to operate across an environment containing far more machine identities and autonomous actions. The objective is to limit the consequences when something goes wrong and maintain the ability to recover quickly, rather than designing around the assumption that every failure can be prevented. 

The speakers described AI as an unusually cross-domain attack surface because it connects identity and access management, endpoints, cloud services, data stores, third-party risk and business applications. Defending that environment requires visibility across those domains and processes capable of responding at comparable speed to AI-enabled attacks. Permission limits, containment, rollback mechanisms and human intervention can then work together to prevent an unexpected action from becoming a wider business incident. 

Red lines should reflect risk and control maturity

Greater autonomy also raises the question of where organizations should draw boundaries around AI use. There may be decisions for which the potential consequences make full automation inappropriate, but those boundaries will differ according to industry, business context and an organization's ability to control the associated risk.

Jain argued that both risk appetite and security maturity should inform those decisions. 

An organization with strong identity controls, effective vulnerability management and mature monitoring may be able to give an AI system more authority than one carrying significant gaps across the same areas. This makes AI posture and broader cybersecurity posture closely connected: where foundational controls are weak, the boundary for acceptable autonomy may need to be more conservative. 

Boards and executive leaders need to understand where AI errors or hallucinations could create material business consequences, what safeguards can reduce that exposure and what level of residual risk the organization is prepared to accept. Where the risk cannot be reduced sufficiently, continued human involvement or a decision not to automate may be the appropriate response. 

Trust must be earned continuously

Mehta summarized the objective as “the ability to earn that trust every day.” 

That principle connects many of the challenges emerging around enterprise AI. Trust depends on knowing which identities are operating, understanding the authority they have, continuously observing their behavior and maintaining accountability for the decisions they make. It also depends on resilience: organizations need safeguards that allow them to contain unexpected behavior and recover when controls do not work as intended.

Done well, those security disciplines can support rather than constrain AI adoption. Leaders can move faster when they have confidence in where risk sits, how systems are behaving and what will happen if something goes wrong. As autonomous systems become more deeply embedded across the enterprise, continuous assurance and resilience will increasingly determine how confidently organizations can scale AI while preserving the security and trust their businesses depend on.