Contact us
Request a demo

AI agents and IAM: entitlements don’t disappear, they become a ceiling

ANSSI and NIST recommendations on AI agents converge: the entitlement granted becomes a ceiling, and the decision is made action by action. Why the agent’s mandate links IGA and access control, and where to start.

An AI agent doesn’t just access data. It chooses its tools, chains actions together and can delegate to other agents. Does that mean we need to invent a new IAM?


No. But its center of gravity has to shift. The recommendations published at the end of October 2026 by ANSSI, the French national cybersecurity agency, and by NIST show this, each in its own way.


An entitlement, once granted, doesn’t go away: it becomes a ceiling. The decision itself is now made action by action.


The two texts don’t have the same status:


ANSSI’s guidance, “Securing an agentic AI system”, covers the security of the agentic system as a whole.

NIST’s blog post sets out ten non-normative design tenets, centered on identity, delegation and authorization.


That makes their convergence more significant.


IAM fundamentals remain the foundation

Let’s start with what doesn’t change. An agent must be identified, governed and tied to a human who answers for it.


ANSSI asks organizations to inventory the access and capabilities of an agentic system and to justify them against strict business need. It also requires a named human owner for each security measure. NIST asks for the agent to be verifiably linked to the person or organization it acts for.


These are the basics of the identity lifecycle. An IGA solution such as MY-Identity already applies them to employees and technical accounts alike. But that foundation isn’t enough to govern an agent.


NIST sets a condition: each agent must have its own identity, distinct from those of humans and service accounts. ANSSI is more cautious. It asks for identities dedicated to each environment, without explicitly prescribing one identity per agent.


We side with NIST here. The line with a service account isn’t absolute, but a service account mostly performs actions planned in advance. An agent decides its own actions, and they vary from one run to the next. Running it under a shared account means losing the ability to tie each of its actions back to someone accountable.


The real shift: authorization is decided at the action

IGA was built around one question: who holds which entitlement? With agents, that question is still necessary, but it’s no longer enough.


ANSSI puts it bluntly: “an authorized action can still cause damage.” An agent can make mistakes, be manipulated, or combine permitted operations in dangerous ways. ANSSI therefore asks for control over the action itself: its parameters, its target, its compliance with business rules. It also asks for sequences of actions to be checked.


NIST goes in the same direction. It asks for every tool call to be verified at execution time, based on context and risk.


On one point, the two texts clearly converge: the authorization decision must not belong to the model. Where feasible, ANSSI recommends deterministic filtering, independent of the agent, applying rules the model cannot change. NIST asks for a separation, from the architecture stage, between what the agent attempts to do and what it is allowed to do.


Delegation makes this requirement stronger. According to NIST, the scope of rights should narrow at each level of delegation, and the agent should hold no standing privileges. ANSSI, for its part, recommends short-lived access tokens limited to what is strictly necessary.


Authenticating an agent and issuing it a token is therefore not the same as authorizing each of its actions. The entitlement granted sets the maximum the agent could do. Each action must then answer three questions: on whose behalf, under which mandate, and within what limits?


The mandate: where IGA and access control meet

If entitlements become a ceiling, something must justify each action below that ceiling. For us, that something is the mandate.


The term matters. ANSSI speaks of a “mission”: it asks for it to be documented from the design stage of an agentic system and reviewed at least once a year, and after any significant change. We prefer to speak of a mandate, because the word says more. It implies that a person has mandated the agent to carry out a specific task, which ties the agent to human delegation and to the accountability that comes with it. NIST proposes a similar form, without making it a requirement: mandates signed by the person who initiates the task, describing the expected outcomes and the constraints to respect. These mandates can derive from templates defined by the organization.


For an IGA team, none of this is unfamiliar. A mandate template looks like a role template. A mandate review looks like a recertification. One difference remains: this review must also be triggered when the agent’s capabilities or risk change, not just on a calendar schedule.


The mandate thus becomes the link between two worlds that rarely talked to each other. Identity governance says what is justified and who answers for it. Access control says what is allowed right now. With agents, these two dimensions can no longer be run as separate projects.


There is an organizational stake too. If IAM teams don’t take up this subject, agents’ access decisions will be made elsewhere: in tool configuration, case by case, with no common reference framework.

Matching intent against the mandate, at every tool call

A mandate is only worth something if it can be enforced. At every tool call, the agent expresses an intent at a specific moment. The question becomes: does that intent stay within the mandate approved for this agent and this session?


That is the role of a PDP (Policy Decision Point), the authorization decision engine. It doesn’t run inside the agent: it makes the decision, and an enforcement point, the Agent Gateway sitting between the agent and its tools, blocks the call if it is denied. This separation answers what ANSSI and NIST both ask for: the decision doesn’t belong to the model.


We are considering four capabilities for this engine:


  • Static policies: readable rules, written in a console or through an API.
  • A permission model: roles and permissions for agents, based on the owner, the mandate and the human the agent acts for.
  • Mandate drift detection: each call is compared with the approved mandate, and a clear deviation is denied.
  • A risk score: based on context, data sensitivity and behavior over time, always explained by its factors.

The first two capabilities build on what IAM already knows how to do. The last two are the real novelty: the decision no longer relies only on rules written by humans, but also on signals computed at every call. The same risk engine serves humans and agents.


Memority is carrying out innovation work on AI agent governance, in this same spirit of continuity with IAM. It covers in particular automated agent inventory, dynamic delegation and revocation, and a PDP built on open standards (OpenID AuthZEN) to integrate with Agent Gateways. Its goal: to compare every tool call with the agent’s approved mandate.


Par où commencer ? 4 actions pour les équipes IAM


  • Recenser les agents en service et identifier ceux qui utilisent des comptes partagés ou l’identité d’un utilisateur.
  • Rattacher chaque agent à un responsable et à un mandat documenté, revue au moins une fois par an et à chaque modification importante.
  • Privilégier les accès temporaires et les jetons de courte durée, limités au strict nécessaire.
  • Vérifier la traçabilité des demandes, des autorisations et des actions, dans des journaux hors de portée de l’agent.

Les recommandations de l’ANSSI et du NIST ne rendent pas l’IAM obsolète. Elles lui demandent de savoir non seulement ce qu’une identité a le droit de faire, mais aussi ce qu’elle fait, pour qui et dans le cadre de quel mandat.

Sécuriser les identités de vos agents IA

Un agent IA dispose de comptes, de jetons et de droits sur vos applications et vos données. Memority permet de les inventorier, de limiter leurs permissions au strict nécessaire et de tracer chacune de leurs actions.


Découvrir le cas d’usage Agents IA

Duo-working-IAM

Published by

Léa Zerah

Product Lead Gouvernance

Léa Zerah is Product Lead for AI Agent Governance at Memority, where she designs the capabilities that make it possible to identify AI agents, give each one its own identity, and control their rights and actions. She graduated from Télécom Paris and holds a master's degree in cybersecurity and data science / AI, which gives her dual expertise in information systems security and artificial intelligence. She spent three years as Deputy CISO at Memority, where she helped define and strengthen the company's security posture, led risk management, and ensured compliance with the most demanding standards. She now applies her security expertise to a product vision, enabling organizations to innovate with AI with confidence.

Recent articles

Personne utilisant un ordinateur portable devant une interface d'assistant IA, illustrant les agents IA et la gouvernance IAM

AI agents: from answering to acting, why IAM must govern tools and identities

Calendrier

November 19, 2026

Welcome to Le Labo IAM4AI, Memority's series dedicated to innovations at the crossroads of artificial intelligence and IAM. Our experts decipher emerging technologies and explore the new challenges of identity and access management.

IAM at a turning point: toward more visible, intelligent and dynamic identity management

Calendrier

November 17, 2026

IAM must now govern all identities, including service accounts and AI agents. Discover the three key evolutions (Visible, Intelligent, Dynamic) and why IAM is converging with the SOC to address Zero Trust. Alexis de Calan explains it all in video.

Assises 2026: digital identities at the heart of conversations

Calendrier

November 9, 2026

Compliance, AI agents, transformation: what our conversations at the Assises de la cybersécurité 2026 reveal about the IAM priorities of IT and cybersecurity decision-makers.