UNAUTHORIZED artificial intelligence tools are creating cybersecurity risks for enterprises across Asia-Pacific as employees connect AI platforms to corporate systems and data without adequate oversight, according to leading cybersecurity companies.
Keeper Security’s Identity Security at Machine Speed report found that 53 percent of surveyed cybersecurity decision-makers in Asia-Pacific identified the management of AI-related non-human identities as a major gap in identity governance, compared with 43 percent globally. Another 47 percent cited insufficient visibility into employees’ use of AI tools as a significant governance gap.
The findings highlight a risk that extends beyond employees entering confidential information into public AI chatbots. When AI tools connect to enterprise applications, databases and cloud services, they may acquire credentials and permissions that allow them to access information and perform tasks outside established security controls.
“When teams link AI assistants to enterprise calendars, cloud repositories and CRM platforms to accelerate everyday work, they are not just using an unapproved application — they are granting persistent credentials, API keys and elevated data access to machine actors,” Takanori Nishiyama, Keeper Security’s senior vice president for Asia-Pacific and Japan country manager, wrote in a commentary.
The issue is becoming more pressing as enterprises adopt agentic AI, which enables software systems to pursue assigned objectives, access information and perform actions with varying degrees of human supervision.
SailPoint’s Horizons of Identity Security 2026–2027 report found that 79 percent of surveyed organizations were running AI agents in production, but only 2 percent had deployed identity-security tools specifically designed to govern and secure them.
Based on a survey of 340 senior identity, information technology, cybersecurity and risk leaders across North America, Europe, Asia-Pacific and Latin America, the report points to a gap between AI adoption and the security infrastructure needed to control access.
What is shadow AI?
Shadow AI refers to employees’ use of AI tools outside an organization’s approved policies, security controls or governance processes. It extends the concept of shadow IT, which involves software and cloud services adopted without formal IT approval.
Employees may use personal AI accounts because they prefer a particular tool, find approved alternatives difficult to use or want to complete tasks more quickly. This can include uploading documents to public chatbots, processing business information through unapproved AI assistants or connecting external AI services to corporate applications.
Information submitted to an external AI service may be exposed outside the organization, depending on the provider’s data-handling practices and contractual protections. A more serious risk can arise when an AI tool connects to internal systems and receives credentials that allow it to retrieve sensitive information or perform actions on a user’s behalf.
These credentials can include API keys, access tokens, certificates and service-account credentials. They are examples of non-human identities, or NHIs, which applications, machines, automated processes and AI agents use to authenticate and access resources.
Non-human identities are not inherently insecure. The risk arises when organizations cannot identify them, establish ownership, restrict permissions, monitor activities or revoke access when it is no longer required.
An AI agent may execute tasks continuously and interact with multiple systems in rapid succession. If it has excessive permissions or retains credentials indefinitely, a compromised or misconfigured agent could expose information or perform actions beyond its assigned task.
Nishiyama cited IBM’s 2026 Cost of a Data Breach Report, saying 43 percent of breached organizations had experienced an incident involving shadow AI, up from 20 percent a year earlier. He also cited an average breach cost of US$5.39 million. These figures should be checked against the original IBM report before publication.
Adoption outpaces security controls
SailPoint’s report found that 85 percent of surveyed organizations relied on legacy, general-purpose identity tools or platform-native controls to manage access rather than tools specifically designed to secure AI agents. About 60 percent remained in the two lowest levels of SailPoint’s five-tier identity-security maturity model, while less than 1 percent had reached the highest level.
The report also found that 54 percent of organizations were at the lowest maturity level for agent identity security. Among respondents, 35 percent cited data leakage and exposure as a leading AI-agent risk, while 30 percent identified excessive privileges and access-control problems.
These risks are not limited to unauthorized AI tools. Even an approved agent can create security problems if it has more permissions than necessary, retains access after completing a task or operates without sufficient monitoring.
In an interview with The Manila Times on Oct. 20, Eric Kong of SailPoint said AI agents had introduced another category of identity that enterprises needed to manage.
“They gain access to data and systems through credentials. This opens up a whole new class of non-human identity credentials, including tokens, keys and certificates that allow agents to access data and systems,” Kong said.
He said organizations first needed to discover all agents and non-human identities across their environments, including credentials that security teams might not know exist.
“Those credentials are keys to the kingdom. It’s as though all your passwords and keys are lying on the floor, with nobody seeing them,” he said.
Kong outlined three elements of SailPoint’s approach to agent security: discovery; governance; and protection, response and remediation. These involve identifying and registering agents, linking them to accountable human owners, restricting access and monitoring activities for potentially harmful behavior.
Accountability, least privilege
Both companies emphasize the need to apply access controls to AI agents and other machine identities while maintaining continuous oversight of their activities.
Kong said every agent and its associated non-human identities should be linked to an accountable employee or owner. This allows organizations to trace an agent’s actions to the person or process that authorized them and establish responsibility when it performs an unintended task.
“The number one priority for every enterprise should be establishing a link between an agent or non-human identity and an accountable human employee,” he said.
SailPoint advocates just-in-time access, under which an agent receives only the permissions needed for a specific task and for a limited period. Access is revoked when the task is completed, reducing the risk of persistent access to sensitive systems. This supports zero standing privilege, which avoids leaving agents with permanent administrative or other elevated permissions.
Keeper Security similarly recommends discovering AI tools and integrations, rotating credentials, reviewing access, monitoring privileged sessions and removing unnecessary permissions. These measures align with zero-trust principles, which require organizations to verify access requests and grant only the permissions necessary for a task.
Organizations must also monitor agents for unexpected activity and establish procedures to suspend access when serious risks emerge. Depending on the system and potential operational impact, disabling an agent may require automated intervention or human approval.
The objective is not necessarily to prohibit AI adoption but to ensure its use remains visible, accountable and governed. Blocking unauthorized tools without addressing why employees use them may simply push activity further beyond the organization’s visibility.
As AI agents become more deeply integrated into business operations, security teams must identify the credentials behind each connection, establish clear ownership and enforce controls throughout an agent’s lifecycle. The challenge is to ensure AI adoption does not outpace an organization’s ability to determine which systems can access its data, what they are permitted to do and who is accountable for their actions.
