The focus on agentic AI and machine identity at the Gartner Tokyo Security Summit reflects a growing concern for security leaders: autonomous software is becoming an active participant in enterprise systems. AI agents can interpret instructions, select tools and perform actions, but their access must be attributable, constrained and reviewable. That makes machine identity central to the safe adoption of agentic systems.
The problem: agents act with borrowed authority
Many early AI deployments rely on a human user’s session, an application-wide API key or a service account with broad permissions. These patterns are convenient for experimentation, but they obscure who or what actually performed an action. When an agent creates a record, changes a configuration or retrieves sensitive data, logs may show the initiating employee without capturing the agent’s decision path.
This weakens least privilege and complicates investigations. An agent may also chain several tools together, creating an effective privilege set larger than any single integration suggests. Without an independent identity, policy enforcement remains dependent on prompts and application logic that can change quickly.
Machine identity must include behavioural context
Identity for an AI agent should describe more than a name and credential. It should include the agent’s owner, purpose, model or workflow version, approved tools, data boundaries and operating environment. Runtime signals matter as well: an agent requesting an unusual volume of data or acting outside its normal workflow may require additional verification.
This is where NHI security connects identity governance with detection. Static inventories identify what exists; runtime controls show what an identity is doing. Both are necessary to establish whether an automated action is legitimate.
Governance patterns for Agentic Identity
Security teams can begin by assigning every production agent a distinct identity and owner. Permissions should be scoped to individual tasks, with short-lived credentials and explicit approval for sensitive operations. Tool calls should be logged with the agent identity, human sponsor, policy decision and resulting action.
These controls need to be integrated into development pipelines rather than added after deployment. Model changes, new tools and altered prompts can change an agent’s effective behaviour and should trigger review. The summit’s emphasis on machine identity signals that AI security is moving beyond guardrails toward accountable, policy-driven execution.