This document outlines the strategic evolution of the AI SAFE² framework from a conceptual foundation to a production-oriented governance and runtime-enforcement system.
Documentation updated: 2026-10-02. The existing organizational Sovereignty Score is clarified; TCP and Evidence Assurance v1.0 are introduced as CSI methods for use and feedback. Follow the AISM assessment guide for the Technology Card workflow and worked example.
The AISM Technology Contribution Profile separates organizational maturity, scoped artifact contribution, and claim-level assurance. The change includes a reusable card, validation/incident practices, nine source-attributed adoption candidates, and repository-wide navigation and compatibility review. It is not a framework/CLI release, certification program, scoring migration, new control, or completed runtime integration. Existing grades, formulas, versions, generated control data, and frozen Challenge 001 are unchanged.
| Version | Released | Framework Controls | Frameworks | Primary focus |
|---|---|---|---|---|
| v3.1 | August 2026 | 161 | 32 | MCP 2026-07-28, protocol-independent governance, three enforcement planes |
| v3.0 | April 2026 | 161 | 32 | Swarm governance, production evidence, CP.1 through CP.10 |
| v2.1 | November 2025 | 128 | 14 | Agentic and distributed governance |
| v2.0 | October 2025 | 99 | Core | Enterprise operations |
| v1.0 | June 2025 | 10 | Initial | Foundational concepts |
Released: August 2026
Core concept: Governance must survive protocol change.
Version 3.1 responds to a structural change in the Model Context Protocol. MCP 2026-07-28 makes the core protocol substantially more stateless, removes the required initialize/initialized handshake and Mcp-Session-Id, adds optional server/discover, introduces cacheable catalogs and protocol extensions, and strengthens the authorization model.
The security implication is larger than a version bump. A governance framework cannot safely bind identity, authorization, memory, or evidence to protocol-owned lifecycle concepts that may disappear in the next protocol revision. AI SAFE² v3.1 therefore moves the governance boundary to framework-owned constructs.
A CP.5 profile MUST NOT bind a control to a construct owned by the protocol it profiles.
The canonical governance objects are now verified principals, capability grants, delegation chains, provenance baselines, state handles, trust-establishment events, policy decisions, and independently reconstructable evidence.
v3.1 makes three distinct enforcement planes explicit:
| Plane | Traffic | Primary governance concern |
|---|---|---|
| North-south | Agent to model provider | Content, policy, consumption, and spend |
| East-west | Agent to agent | Identity, delegation, lineage, and authority |
| Agent-to-tool | Agent to MCP server or tool | Tool authorization, provenance, output trust, and economic amplification |
The three planes remain separate enforcement points but use a common policy vocabulary, principal model, and evidence model.
The MCP profile expands from 13 to 19 sub-controls while the overall AI SAFE² framework remains 161 controls.
The six new MCP profile controls are:
| Control | Focus |
|---|---|
| MCP-14 | Extension capability negotiation |
| MCP-15 | Header/body assertion integrity |
| MCP-16 | State-handle binding and lifecycle |
| MCP-17 | MRTR round-trip integrity and replay resistance |
| MCP-18 | Catalog-cache integrity and provenance revalidation |
| MCP-19 | Authorization-chain integrity, intended resource/audience validation, and SSRF boundaries |
These are profile-level controls under CP.5. They do not change the top-level 161-control count.
AI SAFE² specifies what must be governed and enforced. NEXUS provides Cyber Strategy Institute’s first-party reference implementation across the agent-to-agent and agent-to-tool planes. Organizations may use NEXUS or another implementation that demonstrably satisfies the applicable controls and required evidence.
The v3.1 NEXUS MCP adapter defines a fail-closed interface contract. Unimplemented methods raise rather than silently allowing traffic. It is an implementation contract, not a claim that the adapter is production-ready.
MCP 2025-11-25 remains supported as a legacy binding for twelve months. Legacy Mcp-Session-Id values are treated as principal-scoped state handles, not identity or authorization boundaries.
server/discover is optional in MCP 2026-07-28 and is not required for AI SAFE² conformance.
Opaque static bearer tokens do not contain an audience claim. They therefore do not, by themselves, establish MCP-19 audience-validation conformance. A deployment must validate the intended resource or provide equivalent independently evidenced resource binding before claiming the control.
Challenge 001 remains falsification-first. The framework and its reference implementation are subjects of the experiment, not sources of their own proof. v3.1 further requires that claims be scoped to the enforcement planes actually exercised. Successful east-west governance does not automatically validate the MCP/tool plane or north-south provider plane.
Material normative changes to the framework, protocol profile, implementation, policy, or grader after preregistration require a new preregistration version before confirmatory evidence is pooled.
Released: April 2026
Core concept: The Governance OS for Autonomous AI.
Version 3.0 shifted AI SAFE² from framework coverage toward engineering certainty. It introduced the Cross-Pillar Governance Layer, CP.1 through CP.10, and expanded the framework to 161 controls.
Key v3.0 additions included:
Control count: 151 pillar controls plus 10 cross-pillar governance controls = 161 total controls.
v3.1 does not replace those capabilities. It hardens the protocol and enforcement semantics layered on top of them.
Released: November 2025
Version 2.1 expanded the framework to 128 controls and introduced deeper non-human identity, swarm, memory, and distributed-agent governance.
Released: October 2025
Version 2.0 expanded AI SAFE² to 99 controls with enterprise governance and NIST/ISO mapping.
Released: June 2025
Version 1.0 established the original ten-topic foundation that later evolved into the five-pillar model.