Cybersecurity leaders are warning that the arrival of AI agents is not merely changing how work gets done. It is also changing who, or what, gets access to an organisation’s most sensitive systems. As enterprises race to deploy agentic tools that can act on behalf of users, a new class of identity, privilege and exposure problem is emerging alongside the productivity gains.

According to TechTalkThai, experts at a recent discussion argued that organisations can no longer treat AI as a passive content generator. Once an AI system is allowed to take action, it must be granted credentials, permissions and connectivity, which means security teams have to think about it as a privileged identity rather than a simple application feature.

AI agents are creating a wider and stranger attack surface

The core concern is not only that AI can help defenders work faster. It is that the same technology can help attackers move faster too. AI agents can operate continuously, map networks, search for weak points and probe systems around the clock, creating pressure on teams that still rely on periodic testing and manual oversight.

TechTalkThai reported that Matt Hartman, former acting head of cyber at CISA and now chief strategy officer at Merlin Group, said the shift from content generation to action means AI agents must be given access to systems and data. His view was that every AI agent should be handled like a privileged identity because it can reach places that ordinary users never need to touch.

That creates a second problem for enterprises: the rise of non-human accounts. These machine identities can grow quickly, be hard to inventory and often slip past legacy policy frameworks that were built for people, not autonomous software. In practice, that means the security perimeter is no longer just about employees, contractors and partners. It now includes software entities that can act at machine speed and with broad permissions.

Phishing, social engineering and trust signals are under strain

The external threat picture is also becoming more difficult to read. AI-assisted phishing and social engineering can be tailored with a level of precision that undermines the usual clues staff are trained to spot. Traditional trust signals, once used to judge whether a message or request seemed legitimate, are proving less reliable when attackers can automate personalisation at scale.

TechTalkThai said Hartman argued that defenders need to shift towards zero-trust architecture, phishing-resistant authentication and behavioural analytics. That combination reflects a broader industry recognition that identity assurance cannot depend on a single check or on the assumption that a request looks familiar. Instead, organisations need layered controls that verify context, device posture and user behaviour before granting access.

The implication for Indian enterprises is significant. As AI adoption expands across banking, manufacturing, technology services and public-sector workflows, the number of digital identities inside an organisation will rise sharply. The challenge is not only issuing access, but also continuously validating whether that access is still justified and whether the system using it is behaving as expected.

From annual pen tests to continuous automated red teaming

The security testing model is also being forced to evolve. Rob Joyce, former cyber executive at the National Security Agency, was cited by TechTalkThai as saying that enterprise systems will be red-teamed whether organisations pay for it or not. The logic is straightforward: adversaries are already probing for weaknesses, and AI agents can do that work continuously without fatigue.

That is why automated red teaming is gaining attention. Rather than relying only on a once-a-year or once-a-quarter penetration test to satisfy compliance requirements, security teams are being pushed towards continuous, machine-driven simulation of attacker behaviour. The objective is to find misconfigurations, exposed pathways and business logic flaws before real adversaries do.

Jay Bavisi, founder of EC-Council, was quoted by TechTalkThai as saying that annual or quarterly testing done mainly for compliance is outdated. His argument was that hackers are not constrained by calendar cycles or narrow scopes, because they use automation to search across the enterprise all the time. In that environment, testing has to become more persistent and more realistic.

Human testers are not disappearing, but their role is changing

Even as automation takes on more of the repetitive probing, human penetration testers are unlikely to become redundant. The more likely outcome is that their work shifts towards higher-value assessments that machines cannot easily replicate. That includes stress-testing large language models, studying the behaviour of AI agents and measuring what happens when an AI system is compromised or begins to act incorrectly.

This is where the story moves beyond traditional cybersecurity into operational risk. If an AI agent is allowed to call tools, access data and trigger workflows, a failure is no longer just a technical issue. It can become a business issue, affecting customer service, financial controls, compliance obligations or internal decision-making. Security teams therefore need to ask not only whether a model can be attacked, but also what happens to the organisation if the model makes the wrong move.

For enterprises, the practical takeaway is clear. AI deployment cannot be separated from identity governance, access control and continuous testing. The same systems that promise efficiency can also expand exposure if they are not wrapped in stricter controls. In that sense, automated red teaming is less a niche security product than a response to a structural change in how enterprise software now behaves.

As TechTalkThai noted, the emerging lesson is blunt: if organisations do not use AI to attack their own systems, adversaries will do it for them. The next phase of cybersecurity will be shaped by how quickly defenders can match machine-driven offence with machine-driven defence, while still keeping human expertise in the loop for judgement, context and accountability.

Join the community of 2M+ industry professionals.

Subscribe to Newsletter to get latest insights & analysis in your inbox.

All about ETCISO industry right on your smartphone!