Evidence library / Communities and Working groups

CoSAI Workstream 4: agentic security

Vendor-neutral architecture guidance for securing agentic systems, from an OASIS Open Project. These publications supply design and assurance guidance. This note reports selected themes and AAES Govern evidence; it is not a CoSAI review or certification.

Last reviewed:

This page describes AAES Govern’s controls. AAES Operate has private-work permissions, operational reviews and recorded-spend limits; it does not include Govern. About AAES Govern

The Coalition for Secure AI (CoSAI) publishes guidance on securing systems that use tools and take actions. Its Workstream 4 publications informed selected AAES design decisions. This page explains those decisions and the evidence available to inspect; it does not claim CoSAI review or endorsement.

These pages describe the product. They are not a certification or a legal opinion.

Reading this note

The workstream's own publications are authoritative for their content; dates are cited because the workstream is active and later revisions may change recommendations. AAES uses these publications as design inputs. Nothing here implies CoSAI review or endorsement of AAES.

Instrument identity

Title
CoSAI Workstream 4 (agentic security): the CoSAI Principles, MCP Security v2.0 (published 12 August 2026, announced 25 September 2026), and Agentic Identity and Access Management (approved 20 March 2026)
Issuing body
Coalition for Secure AI (CoSAI), an OASIS Open Project
Instrument type
Vendor-neutral architecture guidance (not a standard, not a certification program)
Official sources
Coalition for Secure AI, the approved Agentic Identity and Access Management paper, the MCP Security v2.0 announcement, the CoSAI OASIS repositories, and the Workstream 4 mailing list (public archives).

Question

Which AAES design decisions did the workstream's publications bear on, and can a reviewer inspect the results?

The January 2026 MCP security paper bears on bypass prevention on every in-scope path (AARM R1.02) and least privilege at the tool boundary (R7); the agentic IAM paper (March 2026) bears on identity binding (R6) and the delegation-chain narrowing rules (R6.06); the CoSAI Principles line up with the scope discipline AARM's S.01 asks for. Review AAES's connector and delegation designs against these publications. Row numbers refer to the AARM/CS 1 draft catalog AAES maps, whose numbering differs from public AARM v1.0.

Selected contribution

The following controls sit beside the workstream's published themes. Their scope is the action layer, not the workstream's catalog as a whole.

Selected CoSAI WS4 themes and AAES Govern control boundaries
WS4 publication themeWhat a reviewer can examine in AAES GovernBoundary
MCP Security (January 2026 themes; v2.0 published August 2026)The connector enforcement contract: an MCP-facing connector must declare its custody model, and a pass-through capability is refused at the connector gate rather than silently governed. Least privilege at the tool boundary: the callable capability list is filtered by access, while the requestable catalog can show unheld capabilities.The contract governs connectors registered with AAES. Activity outside AAES Govern is visible only where telemetry is supplied.
Agentic identity and access management (March 2026)Identity binding: every registered agent has at least one accountable manager, ultimately answerable to a person. Delegation chains: per-hop narrowing is verified, so a delegated grant can only shrink in scope, never widen.Recorded identity and delegation shape do not establish that an underlying human decision was correct.
CoSAI Principles (scope discipline)The internal claim document excludes alternate credential paths and pass-through custody, and records the remaining organizational requirements.The claim document is internal and unsubmitted; it is a claim document, not a conformance claim.

Preconditions and gaps

  • Version-specific assessment remains open. The IAM paper was approved on 20 March 2026. MCP Security v2.0 was published on 12 August 2026, announced on 25 September 2026, and adds deployment-specific assurance profiles and evidence expectations. The table identifies selected themes; AAES has not published a complete assessment against the approved IAM paper or the v2.0 profiles.
  • Design input is not review. CoSAI has not reviewed AAES, and these mappings are AAES's own reading.
  • Credential-path control is required. Activity outside AAES Govern is visible only where telemetry is supplied. Observation is not enforcement.
  • No production adoption as of this review. Deployments are self-hosted or hosted by AAES.

For implementation and testing details beyond this note, use Security and implementation posture and the evaluation page. Results from a client's test tenant are produced during a scoped evaluation.

Client responsibility

The organization makes its own architecture decisions. The workstream's publications are design inputs to that work, not obligations AAES discharges for the client. The client remains the regulated entity.

Choose the tool actions to test, identify who holds their credentials, and list the questions the decision records should answer.

Free workflow assessment