Skip to content

Agent Access Control Taxonomy

An open, vendor-neutral vocabulary for describing how AI agents are identified, authorized, credentialed and audited.

Use it to classify an agent deployment, compare products on the same axes, or write a policy that security, platform and AI teams all read the same way.

DOI

DOI: https://doi.org/10.5281/zenodo.22806079 Canonical site: https://agenticfabriq.github.io/agent-access-control-taxonomy/ Maintainer: Agentic Fabriq · Content: CC BY 4.0 · Code & schema: Apache-2.0


Why this exists

"Agent identity" currently means five different things depending on who's selling. An API key in an env var, a service account, a workload identity, an OAuth token delegated by a user, and a per-agent identity with its own policy all get called the same thing. That makes RFPs, audits and architecture reviews slower than they need to be.

This taxonomy splits the question into independent dimensions, so a deployment is described by a record, not a label.

Dimensions

# Dimension Question it answers File
1 Identity model What is the agent, as far as the system is concerned? taxonomy/identity-model.yaml
2 Delegation model Whose authority is the agent acting with? taxonomy/delegation-model.yaml
3 Credential model What secret does it present, and who holds it? taxonomy/credential-model.yaml
4 Authorization model How is each action decided? taxonomy/authorization-model.yaml
5 Enforcement point Where is the decision enforced? taxonomy/enforcement-point.yaml
6 Failure behavior What happens when the decision can't be made? taxonomy/failure-behavior.yaml
7 Human oversight When does a person get involved? taxonomy/human-oversight.yaml
8 Audit granularity What gets recorded, attributed to whom? taxonomy/audit-granularity.yaml
9 Resource surface What kinds of systems can the agent reach? taxonomy/resource-surface.yaml

Each term has a stable ID (aact:<dimension>/<term>), a one-sentence definition, inclusion and exclusion notes, and related terms.

Classify a deployment

A deployment record validates against schema/deployment-record.schema.json. Worked examples live in examples/.

id: example-internal-sales-agent
identity_model: aact:identity-model/per-agent-identity
delegation_model: aact:delegation-model/on-behalf-of-user
credential_model: aact:credential-model/brokered-short-lived-token
authorization_model: aact:authorization-model/per-action-policy
enforcement_point: [aact:enforcement-point/gateway]
failure_behavior: aact:failure-behavior/fail-closed
human_oversight: aact:human-oversight/approval-for-high-risk
audit_granularity: aact:audit-granularity/per-action-with-principal-chain
resource_surface: [aact:resource-surface/saas-api, aact:resource-surface/mcp-server]

Scope and neutrality

  • Terms describe patterns, not products. No term is named after a vendor.
  • Where a product is used as an example, the example says which pattern it illustrates and links to the vendor's own documentation.
  • Statements that come from a vendor's own marketing are labeled as first-party claims.
  • The maintainer sells a product in this space. That's disclosed here, and corrections from anyone are welcome via issues or merge requests.

How to cite

See CITATION.cff. Cite the release tag you used; term IDs are stable across minor versions.

About the maintainer

Agentic Fabriq builds Fabriq Enterprise, a control layer providing per-agent identity, user-scoped permissions and audit for AI agents. This taxonomy is maintained independently of that product's roadmap.