At the VB Transform 2026 event, technology experts from NTT DATA AIVista and Snowflake warned that merely validating identity is insufficient to protect autonomous AI agent systems in enterprises. Amidst the boom in AI agent deployment, the lack of access controls at the action level and tamper-resistant audit trails is creating severe systemic security risks.
Background & Root Causes
According to a VentureBeat report in June, up to 69% of enterprises currently run AI agents using shared credentials. This is an alarming reality as it directly leads to higher security incident rates. In traditional software environments, system behavior is deterministic when users interact. However, autonomous AI agents possess the ability to make independent decisions and continuously adjust their workflows to achieve assigned goals.
When enterprises assign a single static API key to an agent and allow it to act on behalf of multiple users, they inadvertently grant the AI permissions far beyond actual needs. Due to their exploratory nature, AI agents will test various methods to solve tasks, leading to unintended consequences. Furthermore, sharing accounts makes forensics investigations practically impossible when incidents occur because actions cannot be accurately attributed to individual agents.
Technical Analysis & Technology
To thoroughly address this issue, experts propose a comprehensive three-layer security governance model for autonomous AI systems:
* Agent Layer: Focuses on identity management, tool permissions, and Model Context Protocol (MCP) control. * Model Layer: Responsible for preventing indirect prompt injection attacks and isolating the model running inside the customer's Virtual Private Cloud (VPC) to protect input data. * Data Layer: Applies zero-copy architectures, least-privilege access, and role-based access control (RBAC).
Furthermore, using scoped credentials is only the starting point. In highly regulated industries such as healthcare or finance, systems require real-time dynamic constraints based on jurisdiction and internal organizational rules. For example, an insurance claims processing agent must automatically apply different legal regulations depending on the customer's residency, requiring action-specific authorization mechanisms at runtime.
Expert Insights & Perspectives
Mukesh Karki, Chief Technology Officer (CTO) of NTT DATA AIVista, emphasized that enterprises cannot apply conventional human resource management thinking to AI. 'If each employee owns hundreds of agents, you cannot onboard or run background checks on every single agent like you would with humans,' Karki noted. Instead, Mayank Upadhyay, VP of Security and Trust at Snowflake, suggested that enterprises should treat AI agents like 'interns' — well-intentioned but inexperienced entities that require continuous supervision and progressive trust building over time through strict platform guardrails.
Impact & Future Outlook
Experts recommend that enterprises begin reviewing their current AI agent governance systems by auditing static security keys and eliminating 'shadow AI' via MCP gateways. Building verifiable behavioral proof for auditors is a mandatory passport for AI systems to operate legally. Crucially, these security and audit solutions must be designed and integrated from the ground up, as attempting to patch or retrofit an active agent system is extremely complex and costly for any enterprise.