AI agents are the fastest-growing category of non-human identity.

Why AI agents break the existing NHI model

A traditional service account has a well-defined behaviour envelope. You know what it talks to, because a human engineer wrote the integration. You can describe its authorised scope in a sentence. It is deterministic.

An AI agent is a different animal. It:

• Selects tools at runtime from a set exposed via MCP (Model Context Protocol), A2A (Agent-to-Agent protocol), or similar. A human engineer did not pre-wire which tool is called in which order.
• Is prompted by other agents or by chains of inputs, some of which may be adversarial (prompt injection). The identity authenticating the request and the intent behind the request are increasingly decoupled.
• Operates across identity boundaries on behalf of users, on behalf of tenants, and on behalf of itself, sometimes simultaneously.
• Is often overprivileged by default because the scope needed for a generic assistant to be useful is broad, and because scoping agent capabilities narrowly is hard UX.

The governance model has to evolve. Articles in this hub work through what that looks like in practice.

Where the regulatory frontier sits

The EU AI Act (Regulation 2024/1689) is the first major legal framework that directly regulates the behaviour of agentic systems deployed in the EU. High-risk systems, a category that captures many enterprise agent deployments, must satisfy risk management, data governance, technical documentation, human oversight, and accuracy/robustness requirements. General-purpose AI models (and GPAI models with systemic risk) have their own obligation set, with compliance deadlines through 2026 to 2027. For the identity community, the practical implication is: every agent deployed in a regulated industry will need identity, scope, and audit trail that satisfies regulators. The lightweight implementations shipped by 2024's experimental frameworks will not.

See EU AI Act, NHI implications for the mapping.

Breach context

No breach of major public significance has yet been attributed solely to an AI agent's actions. But the Salesloft Drift breach of 2025 (which involved an AI-powered SaaS integration) is the closest we have to a real-world study of how an AI-adjacent NHI can be leveraged at scale. As agents proliferate and their scopes expand, the breach ledger will grow.