AI AGENTS are entering enterprise production environments faster than organizations are developing the identity-security controls needed to govern them, according to SailPoint’s The Horizons of Identity Security 2026–2027 report.
The report, based on research conducted in 2026, surveyed 340 qualified identity-security, information technology, cybersecurity and risk leaders across global enterprises. It found that 79 percent of surveyed organizations were running AI agents in production, while only 2 percent used tools specifically designed for agent identities. Just 4 percent reported having a fully modern identity stack.
The findings highlight a growing challenge for businesses adopting autonomous systems: AI agents can access corporate applications, systems and data, but traditional identity controls may not be designed to manage their permissions and activities at machine speed.
In an exclusive one-on-one interview with The Manila Times on Oct. 20, Eric Kong, managing director for Asean at SailPoint, said the adoption of AI agents had reached an inflection point.
“Fast-forward six to 12 months, and everybody has deployed agents,” Kong said. “It’s happening quickly, but it is consistent with global trends.” The report found that 60 percent of organizations remained in the two lowest levels of its five-tier identity-security maturity model. Forty-one percent were at Horizon 1 and 19 percent at Horizon 2, levels associated with fragmented or largely manual identity processes and limited governance of nonhuman identities.
Identity security beyond employees
SailPoint specializes in identity governance and administration, or IGA, which helps organizations manage who can access particular applications, systems and data, and whether those permissions are appropriate.
Kong said identity governance initially focused on employees before expanding to contractors, service accounts, bots and other machine identities. AI agents add another category because they can act autonomously while using digital identities and credentials to access enterprise resources.
“IGA is an area in which we help organizations establish who within the organization has access to which resources and applications, and ensure that they have access only to the applications, resources and data they need,” he said.
This approach follows the principle of least privilege, under which an identity receives only the access required to perform its assigned tasks. Kong said AI agents need the same discipline because they use credentials, tokens, keys and certificates to access enterprise systems.
“As you can see, from humans to machines, agents and nonhuman identities, the identity problem has exploded,” Kong said.
SailPoint’s report found that only 15 percent of organizations could provision nonhuman access in real time. It also found that 58 percent of respondents considered managing the full credential lifecycle for agents difficult, while 57 percent identified monitoring agent behavior and enforcing policies in real time as difficult.
The problem of shadow AI
Kong said SailPoint’s approach to agentic identity security rests on three pillars: discovery, governance, and protection, response and remediation.
“We believe agentic identity now forms the fabric of any security architecture. You need to weave it into the foundation of your security architecture,” he said.
Discovery means identifying AI agents and nonhuman identities across an organization, including those deployed without formal approval. Kong said some organizations discover thousands or tens of thousands of credentials during assessments, despite initially believing they have few such identities.
“Being able to discover and register every agent and nonhuman identity, or NHI, is critical,” he said.
He also identified shadow AI — the use of artificial intelligence tools outside an organization’s approved policies — as a governance concern.
“Shadow AI is the new shadow IT. It refers to AI being used outside an organization’s sanctioned usage policies,” Kong said.
Employees may use personal accounts or external AI services because they prefer their features or are accustomed to them. The risk arises when they upload company information to those tools, potentially exposing financial, customer or employee data.
“There’s a real need for organizations to bring shadow AI under control, discover all shadow AI usage, block unauthorized usage and educate employees about safe AI practices,” Kong said.
Governance also requires linking every agent or nonhuman identity to an accountable person. Organizations need to know who authorized an agent, what credentials it used, which information it accessed and whether its actions were consistent with its assigned task.
“The number one priority for every enterprise should be establishing a link between an agent or nonhuman identity and an accountable human employee,” Kong said.
He said a traceable record of access and activity allows organizations and auditors to reconstruct an incident and identify the person who authorized the agent and the credentials it used.
Limiting access, tracing responsibility
Kong advocated just-in-time access, in which an agent receives permissions for a specific task and loses them when the task is complete. This reduces reliance on standing privileges — access rights that remain available even when they are not needed.
“Our platform enables real-time, just-in-time access. Agents receive only the access they need for a limited period to perform a task. Once the task is completed, that access is removed,” he said.
Kong described this as zero standing privilege, a way to limit the time and scope of access without preventing an agent from completing its assigned work.
He also warned that AI agents can behave unexpectedly when they encounter obstacles. An agent trying to achieve a goal may attempt alternative approaches, potentially moving beyond its original purpose.
“The ability to continuously monitor agent behavior, detect anomalous activity, identify behavioral drift and catch problems before an agent goes rogue is critical,” Kong said.
A kill switch can disable an agent when a serious problem is detected, but Kong said the decision to make that action automatic or require human approval depends on policy and operational risk. Abruptly disabling an agent performing a critical function could disrupt business operations.
Data sovereignty, regulatory readiness
The SailPoint executive also addressed data sovereignty and the challenges of meeting different regulatory requirements across markets.
Kong said data-sovereignty rules should be interpreted according to the principles they seek to enforce and the types of information that need protection.
“It doesn’t mean that the data sovereignty would apply to all data,” he said.
He said organizations should classify information according to its sensitivity and determine which regulatory requirements apply. A broad requirement to keep all data within a particular jurisdiction could constrain business operations if it does not distinguish between different types of information.
SailPoint’s report found that 57 percent of respondents expressed confidence in their ability to understand and meet regulatory requirements for AI security and governance, while 43 percent said they were highly prepared to produce evidence for an AI-related audit inquiry. These figures measure different aspects of readiness.
Kong said regulatory maturity varies across Southeast Asia, although organizations face common challenges in establishing accountability, managing access and producing evidence for audits. He also described SailPoint’s engagement with regulators, including participation in consultation on AI governance and security in Singapore.
The same report found that 62 percent of respondents reported notable productivity improvements from investments in nonhuman identity security, while 46 percent reported safer, faster AI deployment. These are outcomes reported by survey respondents, not independently measured results.
Identity security must evolve as organizations expand their use of autonomous systems, according to Kong. Businesses need to identify the agents operating in their environments, limit their permissions, monitor their behavior and maintain a clear link between each agent and the person accountable for its actions.
