Cerbos 2026 Benchmark Sets Agent Authorization and Auditability Standards
Summary
Cerbos’s 2026 guide outlines an authorization maturity model, arguing that agent access should preserve human-user permissions and delegation records. It recommends a 90-day program to assess authorization, pilot a central policy layer, and measure progress.
Key points
- The guide says agent authorization should preserve the human principal’s permissions across each delegation hop, rather than rely on broad service-account access.
- It describes Stage 4 maturity as central policy evaluation across the full delegation chain, continuous decision logging, and real-time context-based authorization.
- It argues audit records should identify the human principal, agent and model version, authorizing policy, and delegation scope so actions can be reconstructed.
- The proposed 90-day plan covers assigning an owner, mapping authorization and AI agents, selecting and piloting a central policy layer, and presenting measured results to executives.
- Cerbos authored the guide and promotes its authorization platform as a way to implement the controls it recommends.
Article Details
- Publisher
- Cerbos
- Report Period
- 2026
- Scope
- Authorization maturity, including governance and auditability for AI agents and non-human identities.
- Sample Size
- Based on hundreds of conversations with security leaders, analysts, and regulators; no formal study sample is specified.
- Key Statistics
- The guide presents a 90-day plan for establishing visibility, selecting an authorization layer, and demonstrating an initial approach.
- It recommends documenting authorization enforcement and evidence for the five services handling the most sensitive data.
- It describes end-to-end reconstruction of agent actions across a rolling 30-day window as a working definition of audit-readiness.
- Recommendations
- Inventory production AI agents and autonomous workloads, including their human owners, credentials, reachable systems, and the data classifications of those systems.
- Document the current authorization state, identify fragmented or poorly understood services, and assign a named owner to the authorization program.
- Pilot a central authorization layer on a high-risk service and measure latency, evidence quality, and developer adoption.
- For agents, use per-instance identities and scoped, time-limited credentials; preserve delegation chains and record the human principal in decision logs.
- Test authorization policies against the full delegation chain and verify that decision logs support end-to-end reconstruction.
People
Alex OlivierLead author of the guide.Anna PaykinaContributor to the guide.Emre BaranContributor to the guide.Jonathan CareKuppingerCole contributor whose framing informed vendor questions paraphrased in the guide.Kara SpragueHer published work at HackerOne is cited as an input to the guide.Kim CameronHis 2005 Laws of Identity are cited as a foundation for discussion of justifiable action and agent auditability.Phillip MesserschmidtHis published work is cited as an input to the guide.Reiner MertensCited for making the case at EIC 2026 that authorization standards and tools are mature, while organizational positioning remains a gap.
Vendors
Products
Industries
B2B SaaSB2B SaaS, and on the published work of Phillip Messerschmidt, Reiner Mertens, andfintechacross regulated industries including fintech, healthcare, and government, as well asGovernmentacross regulated industries including fintech, healthcare, and government, as well asHealthcareacross regulated industries including fintech, healthcare, and government, as well as