CoSAI · OASIS Open Project · V1.0
Independent companion to the CoSAI Shared Responsibility Framework. Maintained by Bill Stout. Content CC BY 4.0; code Apache-2.0.
Make responsibility clear before you scale AI.
The CoSAI SRF maps every obligation and control to exactly one accountable party per activity, across each layer and operating model. It complements NIST AI RMF, ISO/IEC 42001, and the EU AI Act by naming who is accountable.
Finish with a signed Accountability Decision Record: name one accountable party (persona) per layer for this deployment, record residual gaps, and capture attestation (sign-off).
Create a Decision Record-
Deploying AI
Name which vendor owes which obligation before an incident, across a multi-vendor stack.
Map accountability -
Building AI
Turn each layer’s accountable party into contract language customers can verify.
See obligations -
Governing AI
Name one accountable party for every control. Find gaps before they become incidents.
Start assessment
Autonomy level (L0 to L5) and override tier (T1 to T5) say how independently a system may act and how it can be stopped. They do not change the accountable party. That party remains one named persona per activity.
SRF Stress Test
Describe your AI deployment.
Get an accountability analysis.
The SRF Stress Test takes a plain-language description of your AI scenario (the model, platform, deployment model, and use case) and builds an analysis prompt you copy into your own chat or model. The site does not run an LLM for you.
Open the SRF Stress Test →How the SRF complements what you already use
Other frameworks define what. SRF assigns accountability.
The SRF sits alongside your existing frameworks and assigns one accountable party per activity. NIST AI RMF defines what governance outcomes to achieve. ISO/IEC 42001 defines how to manage AI within your organization. EU AI Act defines which regulatory obligations apply by risk tier and assigns legal roles (provider, deployer, importer, distributor), but it provides no operational model for who owns each layer when an incident crosses vendor boundaries.
-
NIST AI RMF → CoSAI SRF
What to achieve → Who achieves itNIST AI RMF defines the governance outcomes to achieve: Govern, Map, Measure, Manage. SRF adds which party in a multi-vendor deployment is accountable for each outcome.
-
EU AI Act → CoSAI SRF
Which obligations apply → Who at which layerThe EU AI Act defines which regulatory obligations apply by risk tier. SRF maps those obligations to specific layers and operating models: the implementation detail the regulation intentionally leaves to practitioners.
-
ISO 42001 → CoSAI SRF
How to manage AI → Who manages which partISO/IEC 42001 defines how to manage AI within a single organization's boundary (Clause 5.3 roles; Annex A.10 for third-party and customer relationships). Those clauses leave multi-party cloud AI roles underspecified. SRF names who owns each activity across vendors and layers.
Five layers. One accountable party per activity.
Governance requirements cascade from L1 downward. Exactly one accountable party must own each activity; that party can change with the operating model. Where a matrix cell is shared, create a Decision Record that names the persona. That name is the attestation.
Accountability shifts by operating model. In AI-SaaS, the provider owns L3–L5. In IaaS, you own nearly the entire enforcement path. See the full operating model matrix → How this applies across the AI security lifecycle →
Interactive tools
Browser-based. No account. Most tools stay on your device.
Assessment tools and references built on the framework. They keep state in your browser and export as JSON or PDF. The SRF Stress Test builds a copy-paste prompt for your own model (see that tool’s privacy note).
-
Accountability Decision Record
Name one accountable party per layer for a specific deployment, flag residual gaps, and export a signed one-page record.
-
SRF Stress Test
Describe any AI deployment scenario and get a layered accountability analysis, gap flags, and risk summary.
-
Layer Integration Matrix
Who owns what across all four operating models and five layers.
-
Controls Assessment (AICM)
243 CSA AICM controls mapped to SRF layers, with MITRE ATLAS technique crosswalk.
-
AI Security Controls
Layer-by-layer security controls with OWASP LLM Top 10 and ATLAS adversarial threat mapping.
-
AI Regulation Discovery
Filter regulations and standards by industry, geography, and SRF layer. Export a custom reference set.
-
AI Policy Pyramid
Map governance documents to the SRF layer hierarchy and identify gaps in policy coverage.
-
Official System Prompts
Canonical SRF system instructions for LLMs and AI assistants.
-
Join Conventions
How to join the static JSON catalog, ontology edges, and regulation mapping keys.
For LLMs & agents
Built to be read by machines, not just people.
The whole framework is published as plain text, JSON, and a concept graph with stable IDs. Point an agent at any of these to ground it in the SRF.
-
Site index for LLMs
A concise, linked map of the framework, verticals, tools, and data. The place to start an agent.
-
Whole site, one file
Full site content as a single text file for retrieval and bulk ingestion.
-
Machine-readable data
Human landing for every control schema, layer, persona, and responsibility matrix. Raw index at /data/index.json.
-
Knowledge graph
Concept nodes and typed edges for layers, personas, operating models, controls, and mapped standards.
-
Data index (raw JSON)
Programmatic catalog of every data file with schema descriptions and record counts.
-
Retrieval validation
Simulate RAG ingestion against the knowledge pack. Returns matched chunks, concept hits, and a confidence score.
-
GRC matrix export
The accountability matrix as an OSCAL 1.1.2 catalog and a flat CSV for GRC and spreadsheet tools.