Repository navigation
🟠 AI SAFE² v3.5 — Govern the action. Verify the outcome.
173 controls • 157 child clauses • 53 new research notes • Expanded MCP requirements • A navigable Rosetta Stone
Release navigation:
Why this update matters · Before and after · New controls · Research · Rosetta Stone · Tools · Start here · Validation · Next updates
Explore the repository:
Framework home · Control requirements · Research library · Mapping guide · Migration · Merged PR #433
💡 Why this update matters
An agent can hold valid credentials, follow an approved workflow and still act on the wrong target, exceed delegated limits or report success without achieving the intended result.
AI SAFE² v3.5 expands the framework around those operational failures. It connects governance decisions to the conditions immediately before an action, the consequences after execution and the evidence needed to evaluate both.
The focus spans the full lifecycle:
Discover → Establish trust → Delegate → Authorize the effect → Execute → Verify the outcome → Revoke and recover
For builders, this supplies more precise implementation and acceptance requirements. For defenders, it adds failure cases and evidence expectations. For assessors, it makes unsupported, partial and unknown states visible.
This is a framework-wide upgrade across controls, research, AISM, MCP, mappings, reference tooling and repository navigation.
🔄 Before and after
Comparison baseline: AI SAFE² v3.1.
| Area | Previous baseline | v3.5 delta |
|---|---|---|
| Core catalog | 161 controls | 173 controls, adding seven pillar controls and five cross-pillar controls |
| Requirement depth | Existing control descriptions and supporting guidance | 157 explicit child clauses across 53 reviewed decisions, including two research-disposition clauses |
| Research | Existing research library | 53 additional notes connecting source material, placement decisions, alternatives and acceptance cases |
| MCP | 19 profile control IDs | Same 19 IDs, with 12 additional child requirements |
| Deployment context | Existing platform and implementation guidance | Six explicit deployment profiles covering harnesses, gateways, evaluation, browser/computer use, finance and physical systems |
| Threat modeling | Eleven existing threat families | Three sharper profiles deepen coverage without inflating the top-level taxonomy |
| Organizational assessment | AISM v1 scoring | Formula retained; separate scoped readiness evidence and decision support added |
| Human oversight | UAS v1 | UAS v2, retaining 27 IDs with revised human/fairness acceptance and reciprocal CSF references |
| Standards translation | Distributed historical mappings | Consolidated Rosetta records, source-family index, audit rules and an offline explorer |
| Discovery | Folder browsing and distributed references | Task routes, page maps, module navigation, agent indexes and contextual links |
Counts describe framework artifacts. They do not represent deployed protections or certified compliance.
🧱 Twelve new controls
Each control has explicit requirements, research rationale, failure analysis and acceptance expectations.
| ID | New control | Practical focus |
|---|---|---|
| CP.11 | Trust Lifecycle Governance | Revisit trust at discovery, installation, change, admission, commitment and revocation |
| CP.12 | Final-Effect Authorization | Bind authority to the actual consequence and current conditions |
| CP.13 | Control Evidence Independence | Examine who produced the evidence and whether assurance shares the same failure path |
| CP.14 | Organizational AI Management Governance | Establish accountable ownership, deployment necessity and management review |
| CP.15 | Design-Time Coordination Contract | Define agent roles, coordination, dependencies and permitted effects before deployment |
| S1.8 | Pre-Trust Initialization Boundary | Control startup hooks, helpers, plugins and credential access before trust exists |
| F3.6 | Aggregate Consumable-Authority Ceiling | Bound consumption across concurrent work and delegated descendants |
| F3.7 | Environment Binding and Assertion | Bind authority to the intended environment and target |
| A2.7 | Idempotency and Side-Effect Ledger | Handle retries, duplicate actions and uncertain settlement |
| A2.8 | Verified Outcome Contract | Verify the intended state rather than accepting a completion message |
| A2.9 | Evidence-Based Failure Attribution | Separate plausible explanations from supported causes |
| E5.5 | Declared Success Criteria Gate | Define acceptable outcomes before work begins |
🔬 Deeper requirements and research
The update also strengthens existing controls. New top-level IDs are only part of the change.
The 53 research notes preserve:
- Source-specific findings and limitations.
- Why a requirement belongs in an existing control or needs a standalone owner.
- Alternatives, trade-offs and rejected mechanisms.
- Enforcement boundaries and retained human authority.
- Positive, negative and adversarial acceptance cases.
- Recovery requirements and unresolved evidence.
Coverage includes persistent state, memory validity, inter-agent artifacts, delegated identity, public-facing effects, concurrent budgets, evaluation integrity, human comprehension, source rights and proportional adoption.
For example, P2.T4.4 Data Source Mapping gains three clauses covering source identity, intended purpose, provenance, freshness, declaration state and correction. A recorded observation does not establish ownership, a licence or permission to reuse material.
Read the research library · Inspect the expanded decisions · Read the control contracts
🧭 The Rosetta Stone: from source to evidence
AI SAFE² retains its Rosetta Stone role: helping people translate between external requirements, AI risks, control obligations and implementation evidence.
v3.5 brings together:
| Resource | What it provides |
|---|---|
| 1,202 recovered relationships | Traceable historical source-to-control records |
| 33 external source families | A consolidated starting point for standards and framework discovery |
| 64 discovery entries | Additional routes into the source material |
| 16 audit-rule classes | Checks for incomplete, conflicting or stale mapping records |
| Offline explorer | Source, control, review-status and keyword filtering; record details and JSON export |
Follow a relationship through:
External source → AI SAFE² control → Child clause → Research rationale → Implementation boundary → Validation evidence
Recovered relationships retain their review status. This release claims zero approved external coverage mappings from the recovery alone. Missing evidence remains visible.
Mapping guide · Source-family index · Explorer source
To use the explorer, open dashboard/rosetta.html from a local checkout. GitHub’s file view displays its source.
🛠️ Tools, profiles and agent workflows
Framework lookup and evidence review
The CLI adds three routes into the framework:
safe2 framework query --control CP.12
safe2 framework audit
safe2 framework readiness --helpUse them to inspect requirements, audit mapping records and understand the scoped readiness workflow.
Readiness evaluates declared evidence. It does not grant deployment authority.
MCP and deployment profiles
The MCP profile retains its 19 stable control IDs and gains 12 child requirements. Six deployment profiles make applicability clearer across different execution surfaces.
The MCP knowledge server exposes framework information. Enforcement requires a component that mediates the protected action.
AISM and human oversight
AISM v1 keeps its scoring formula. Three threat profiles sharpen attention to assurance compromise, unsafe authorized execution and human-agent trust exploitation.
UAS v2 retains 27 IDs while revising acceptance interpretation for human understanding, appropriate reliance, intervention, fairness and evidence quality. Reciprocal Cognitive Sovereignty Framework references connect human agency with technical control evidence.
Reference behavior and repairs
- A synthetic SQLite authority fixture exercises exact approvals, current conditions, aggregate budgets, duplicate suppression and descendant revocation.
- NEXUS container build and startup paths are repaired.
- Windows MCP wrapper handling bounds input and handles forwarding failures.
- Built CLI/MCP artifacts include current framework requirements.
CLI guide · MCP profile · Ecosystem adoption
🗺️ Navigation for people and agents
The repository now provides:
- A task-focused root README with linked badges, diagrams and direct entry points.
- Shared navigation across current module READMEs.
- A complete page map and machine-readable discovery records.
- Root and NEXUS agent indexes with repository structure.
- Contextual links among controls, research, mappings, examples and validation.
- Explicit dispositions for historical documents, fixtures and independently versioned components.
NEW, UPDATED and RETAINED labels help readers distinguish substantive requirements from preserved baselines.
Complete page map · Agent entry point · Adoption and validation guide
🚀 Start with your task
| Your task | Start here |
|---|---|
| Build or improve an agent | Adoption and validation guide |
| Find exact requirements | Control contracts |
| Understand why a requirement exists | Research library |
| Translate an external framework | Rosetta Stone guide |
| Assess a project and preserve evidence | SAFE2 CLI |
| Connect an assistant to framework knowledge | MCP knowledge server |
| Understand component adoption gaps | Ecosystem adoption register |
| Upgrade from the earlier framework | Migration guide |
Framework v3.5 and component package versions are tracked separately. Follow the component documentation when installing; this framework merge does not publish a new package.
🧪 Validation and compatibility
The final reviewed PR revision, 2792072, completed 47 successful hosted checks, including CodeQL, container build/health, Python matrices, clean-wheel installation, data contracts, navigation, secret checks and dependency checks.
Local qualification recorded:
| Suite | Result |
|---|---|
| Root suite | 1,170 passed; eight skipped |
| MCP suite | 150 passed; eleven skipped |
| NEXUS SDK | 628 passed |
| Framework/explorer regressions | 30 passed |
Qualification found and repaired advanced-control anchor links, explorer link-protocol handling and Markdown formatting errors.
Review limitations: PR-Agent failed to produce a review because its fallback API had exhausted credits. The separate formal security service could not attach to the Windows checkout. Neither is counted as independent approval. The review personas were structured self-review perspectives used by the same assistant.
Hosted checks · Qualification record
Retained compatibility
- Historical v3.0/v3.1 control and MCP datasets remain available.
- AISM v1 arithmetic is unchanged.
- UAS v1 remains versioned; v2 interpretation requires reassessment.
- MCP profile IDs remain stable.
- Component versions retain their own publication and implementation status.
🛡️ Interpret the evidence correctly
| Artifact | Supported use | Boundary |
|---|---|---|
| Control catalog | Define requirements and acceptance expectations | Does not establish implementation |
| Scanner result | Identify supported static findings | Does not prove runtime enforcement |
| Mapping record | Trace a source relationship and its review state | Does not establish legal compliance |
| Readiness output | Review declared evidence and gaps | Does not authorize deployment |
| Local authority fixture | Test modeled transactional behavior | Does not qualify distributed production effects |
| Test result | Support behavior at the tested revision and boundary | Does not establish independent real-world efficacy |
➡️ Dedicated component updates
The framework establishes the v3.5 requirement baseline. Nine issues track the component work needed to implement and qualify it.
| Workstream | Tracking |
|---|---|
| NEXUS delegated authority and distributed containment | #436 |
| Core Gateway enforcement | #434 |
| MCP enforcement and interoperability | #437 |
| Sovereign runtime series, qualified per stack | #438 |
| Scanner clause-level detection coverage | #435 |
| CLI/MCP distribution | #439 |
| Developer Workflow | #440 |
| Direct-source Rosetta validation | #441 |
| UAS/CSF reciprocal evidence validation | #442 |
Start with shared NEXUS authority contracts, then Gateway/MCP enforcement and individual runtime qualification. Scanner and source-validation work can proceed in parallel.
Dependency order and acceptance criteria
Start with one system, one operating boundary and the evidence needed to verify its behavior.
Explore AI SAFE² · Read the controls · Plan your adoption · Report an issue
Licensing: MIT for code under the repository default; CC BY-SA 4.0 for designated framework/documentation material; Apache-2.0 for NEXUS. Subdirectory and third-party terms apply.