ChatGPT, Claude and other AI assistants no longer just answer questions: they use tools, access data and carry out actions. A major shift that is transforming the security and identity governance challenges organizations face.
From chatbot to AI agent: a new dimension for IAM
In many companies, AI assistants are still managed like standard SaaS applications: single sign-on (SSO), data leak prevention and usage policies. But their capabilities are evolving rapidly and call for new control mechanisms.
ChatGPT, Claude and other assistants can now read files, send messages, create tickets or run code. These actions rely on tools, which are accessible in particular through connectors that bundle several features (toolsets), for example in Google Drive, Slack or Salesforce.
As soon as an assistant can use tools to complete tasks autonomously, it enters an agentic logic. For IAM, this means governing not only its identity and its access, but also the actions it is authorized to perform.
Agentic AI: vendors are accelerating tool integration
The agentic AI ecosystem is evolving rapidly. In November 2024, Anthropic introduced the Model Context Protocol (MCP), which makes it easier to interconnect AI assistants, tools and data sources. In 2025, Mistral AI and OpenAI stepped up in turn with the launch of new agentic capabilities.
This evolution comes with ever deeper integration of tools into AI assistants. Connectors are becoming more accessible, and some features can be enabled by default, depending on the platforms and their administration settings.
For companies, this openness is a tremendous opportunity for automation and productivity. It also raises a key question: how do you stay in control of the tools and data that AI agents can access?
The answer cannot rely solely on the settings offered by each vendor. It requires consistent governance of identities and permissions across the entire information system.
AI agent governance: why is controlling tools not enough?
Permissions tailored to each agent's capabilities
The risk level of an AI agent does not depend solely on the model used, but also on the tools and resources it can access. An assistant that reads documents in read-only mode does not have the same capacity to act as an agent authorized to send emails or modify business data.
Governance must therefore take into account the permissions attached to each tool and the actions the agent is authorized to perform. Control can notably take place at the moment of the tool call, when an intent turns into an action.
MCP gateways are one of the technical answers to this challenge. Placed between agents and tools, they make it possible to centralize certain access controls and to allow or deny calls according to defined policies.
AI agents and MCP gateways: the limits of visibility
MCP gateways make it possible to control the exchanges that go through them. But not all AI assistants necessarily go through these control points. The native connectors of platforms such as ChatGPT or Claude can communicate directly with the applications and services they are linked to.
These platforms offer their own access administration mechanisms. In Claude, for example, administrators can set different levels of permission for connected tools, depending on the available features.
The real challenge is the consistency of access policies. When several AI assistants and agent platforms coexist, permissions can be spread across different administration consoles, with no centralized visibility into all access.
AI agent governance: beyond tool control
Even with an MCP gateway, some AI agent activities escape traditional access control mechanisms. This is notably the case for actions performed directly in their execution environment (harness), such as using a terminal or accessing local files.
Other risks appear when data is processed. A document or web page containing malicious instructions can, for example, trigger a prompt injection attack and influence an agent's behavior.
These specificities call for two complementary approaches: governing access to tools and governing the identities of AI agents themselves. The first makes it possible to control permissions and the resources used. The second provides visibility into agents, their owners, their missions and their lifecycle.
Together, they extend IAM governance to the new uses of agentic AI.
Agentic AI: toward a new approach to IAM governance
The rise of AI agents is changing identity and access management practices. Beyond authentication, companies must now control the tools made available to agents, the data they access and the actions they are authorized to perform.
This calls for complementary governance of agents and tools: identify each agent, define its permissions and control the use of resources, regardless of platforms and vendors.
At Memority, we are convinced that IAM has a central role to play in this evolution. By bringing AI agents into unified identity governance, companies can support innovation while keeping control over access.
