1 | Purpose and audience of the Profile | 0 |
2 | Using the Profile: risk-based adoption, shared responsibility and model lineage limits | 0 |
PO.1.1 | Security requirements for development infrastructure and processes include AI model development (High) | 0 |
PO.1.2 | Policies carry the AI-specific security requirements for the software the organisation builds (High) | 0 |
PO.1.3 | Communicate AI model development security requirements to third-party component providers (Medium) | 0 |
PO.2.1 | Roles and responsibilities across the SDLC cover AI model development security (High) | 0 |
PO.2.2 | Role-based training covers AI model threats, vulnerabilities and mitigations (High) | 0 |
PO.2.3 | Leadership commitment to secure development involving AI models (Medium) | 0 |
PO.3.1 | Plan automated toolchains that secure AI model development (High) | 0 |
PO.3.2 | Deploy, operate and verify the AI security toolchains (High) | 0 |
PO.3.3 | Tools generate artefacts evidencing secure AI development, including dataset attestations (Medium) | 0 |
PO.4-5-unchanged | PO tasks applied unchanged from SSDF 1.1: PO.4.2 (Low) and PO.5.2 (Medium) | 0 |
PO.4.1 | Security check criteria include guardrails across the AI development life cycle (Medium) | 0 |
PO.5.1 | Separate and protect each development environment, including training pipelines and model registries (High) | 0 |
PO.5.3 | Continuous monitoring of execution behaviour in development environments hosting AI models (High, new task) | 0 |
PS.1.1 | Store all code and model elements on least-privilege terms, with reward models and weights kept apart (High) | 0 |
PS.1.2 | Protect training, test, fine-tuning and alignment data from unauthorised access and change (High, new task) | 0 |
PS.1.3 | Protect model weights and configuration parameters (High, new task) | 0 |
PS.2.1 | Provide hashes or signatures for models, components and changes (Medium) | 0 |
PS.3.1 | Archive release files with versioned infrastructure tools, model selection rationale and training process records (Low) | 0 |
PS.3.2 | Provenance data for models, components, derivatives and the data behind them (Medium, modified task) | 0 |
PW.1-2-unchanged | PW tasks applied unchanged from SSDF 1.1: PW.1.2, PW.1.3, PW.2.1, PW.4.2 and PW.9.1 | 0 |
PW.1.1 | Risk modelling that covers AI-specific threats (High) | 0 |
PW.3.1 | Analyse data for poisoning, bias, homogeneity and tampering before use (High, new task) | 0 |
PW.3.2 | Track data provenance where known and record where it is unknown (Medium, new task) | 0 |
PW.3.3 | Include adversarial samples in training and testing data under controlled use (Medium, new task) | 0 |
PW.4.1 | Acquire well-secured components, considering reuse of an existing model (High) | 0 |
PW.4.4 | Verify integrity, provenance and security of acquired models and components before use (High) | 0 |
PW.5.1 | Secure coding practice extended to AI, with careful handling of inputs and outputs (High) | 0 |
PW.6 | Build tool security features, with tool versions captured in provenance (PW.6.1 Low, PW.6.2 Medium) | 0 |
PW.7.1 | Code review and analysis policies cover AI model code (PW.7.1 Low) | 0 |
PW.7.2 | Scan all AI models for malware, vulnerabilities and backdoors (High) | 0 |
PW.8.1 | Code testing policies include AI models and the forms of testing to use (High) | 0 |
PW.8.2 | Test all AI models for vulnerabilities and retest after retraining or new data (High) | 0 |
PW.9.2 | Implement and document secure default settings for model administrators (Medium) | 0 |
RV.1.1 | Gather vulnerability information, including from model input and output monitoring and AI vulnerability databases (High) | 0 |
RV.1.2 | Frequent, largely automated scanning and testing of models, with periodic audits (Medium) | 0 |
RV.1.3 | Vulnerability disclosure and remediation policy covers AI models and informs users of limitations (Medium) | 0 |
RV.2.1 | Analyse each vulnerability, including deep analysis of model input and output deviations (Medium) | 0 |
RV.2.2 | Risk responses that allow for rebuild cost and for stopping or rolling back a model (High) | 0 |
RV.3-unchanged | RV tasks applied unchanged from SSDF 1.1: RV.3.2, RV.3.3 and RV.3.4 (Medium) | 0 |
RV.3.1 | Root cause analysis supported by after-the-fact review of training and alignment data (Medium) | 0 |