Secure development frameworks were written for code, but a generative AI model's security also depends on its training data, its weights and the pipeline that produces them, none of which a code review sees. NIST SP 800-218A, Secure Software Development Practices for Generative AI and Dual-Use Foundation Models: An SSDF Community Profile, extends the Secure Software Development Framework (SSDF, SP 800-218 v1.1) with AI-specific practices, tasks, recommendations and priorities. This guide explains what the profile is, who it is for, how its table is structured, the new and modified tasks, the identity and access requirements around model weights, training pipelines and secrets, how to implement and evidence it, and how it maps to other frameworks. It is written for AppSec and ML platform leads, AI product security teams and buyers assessing AI suppliers. For the base framework, see the NIST Secure Software Development Framework (SSDF, SP 800-218) Guide. Publication status was checked on NIST's site on 27 October 2026.
Key takeaways
- It is a supplement, not a replacement. SP 800-218A adds AI content to SSDF 1.1 and NIST states it should not be used without SP 800-218.
- Six new tasks, one new practice. PW.3 (confirm integrity of training, testing, fine-tuning and aligning data) is new, with tasks PW.3.1 to PW.3.3; PO.5.3, PS.1.2 and PS.1.3 are also new.
- Model weights are treated as crown jewels. Least privilege, indirect-only access, integrity monitoring and, where risk warrants, multiparty authorisation and air-gapped environments (PS.1.1, PS.1.3).
- Deployment and operation are out of scope. The profile covers AI model development; runtime controls for deployed AI systems need other guidance.
- Priorities are a starting point. Each task carries a High, Medium or Low priority that organisations are expected to adjust to their own risk.
What it is and who it applies to
| Attribute | Detail (as stated on NIST's publication page) |
|---|---|
| Identifier | NIST SP 800-218A, DOI 10.6028/NIST.SP.800-218A |
| Status | Final, published 26 July 2024 (initial public draft 29 April 2024) |
| Authors | NIST (Booth, Souppaya, Vassilev, Ogata), CISA (Stanley) and Scarfone Cybersecurity |
| Extends | SP 800-218, SSDF Version 1.1 (final, February 2022) |
| Origin | Written in response to Executive Order 14110 on safe, secure and trustworthy AI, which directed NIST to produce an SSDF companion resource for generative AI and dual-use foundation models |
| Audiences | AI model producers, AI system producers (software that uses a model) and AI system acquirers |
| Out of scope | Deployment and operation of AI systems; most of the data governance and management life cycle (cybersecurity of training and development data stays in scope) |
Executive Order 14110 was revoked on 20 January 2025 by the presidential action "Initial Rescissions of Harmful Executive Orders and Actions". The profile itself remains a final NIST publication, and NIST's SSDF project page continues to present it as the finalised community profile; its value as a practice catalogue does not depend on the order. Organisations should check any contractual or procurement requirement that cites it by name. Like SP 800-218, it is voluntary unless a contract, regulator or customer makes it a requirement.
Who should use it: any organisation that trains, fine-tunes or aligns a model (model producer), builds products on third-party models (system producer, where PW.4 on acquired components matters most), or buys AI products and wants supplier questions grounded in a NIST source (acquirer).
Questions acquirers can ask suppliers, grounded in the profile:
- Who and what can read or export production model weights, and is direct human access blocked (PS.1.1, PS.1.3)?
- How are training, fine-tuning and aligning datasets screened for poisoning and tampering, and is provenance recorded (PW.3)?
- Are released models signed or hashed, and does the supplier verify them at load (PS.2.1)?
- How are third-party models and datasets checked before use (PW.4.4)?
- Are model inputs and outputs logged and monitored, and what are the criteria for rolling a model back (RV.1.1, RV.2.2)?
Structure and key requirements
Table 1 of the profile keeps the SSDF's four practice groups (Prepare the Organisation, PO; Protect the Software, PS; Produce Well-Secured Software, PW; Respond to Vulnerabilities, RV) and has five columns: the practice, the task, a priority (High, Medium or Low), AI-specific Recommendations, Considerations and Notes (identified like PO.1.2.R1, PS.1.1.C1, PW.3.2.N1), and informative references to the NIST AI RMF, the OWASP Top 10 for LLM Applications and NIST's adversarial machine learning taxonomy. Tasks that do not exist in SSDF 1.1 are marked "Not part of SSDF 1.1"; others are marked as modified.
New tasks
| Task | Title (summarised) | Priority |
|---|---|---|
| PO.5.3 | Continuously monitor software execution performance and behaviour | High |
| PS.1.2 | Protect all training, testing, fine-tuning and aligning data from unauthorised access and modification | High |
| PS.1.3 | Protect all model weights and configuration parameter data from unauthorised access and modification | High |
| PW.3.1 | Analyse data for signs of data poisoning, bias, homogeneity and tampering before use | High |
| PW.3.2 | Track the provenance, when known, of training, testing, fine-tuning and aligning data | Medium |
| PW.3.3 | Include adversarial samples in training and testing data | Medium |
Notable AI additions to existing tasks
- PO.1.3 (modified): requirements communicated to third parties supplying components should cover AI model development security.
- PO.5 (modified): AI model training environments are named alongside development, build, test and distribution environments; PO.5.1 is High priority.
- PW.1.1: AI-specific risk modelling, with examples including training data poisoning, malicious content in inputs and outputs, adversarial prompt denial of service and theft of model weights; a consideration to check that models are not making significant security decisions without a human in the loop.
- PS.2.1: provide cryptographic hashes or digital signatures for models and their components, and sign model changes. See the Model Signing and OpenSSF Model Signing Guide.
- PW.4.4 (High): verify the integrity, provenance and security of acquired models and components (datasets, reward models, adaptation layers, configuration parameters) before use, and scan and test them for vulnerabilities and malicious content.
- PW.8: include unit, integration, penetration, red team and adversarial testing, and retest after retraining.
- RV.1.1 (High): log, monitor and analyse all model inputs and outputs, and tell users how to report security and performance issues.
- RV.2.2 (High): remediation plans should account for the time and cost of rebuilding models, with criteria for when to stop using a model and when to roll back to a previous version.
Identity and NHI requirements
The profile does not use the term non-human identity, but most of its access requirements land on machine identities: training pipeline service accounts, model registry credentials, CI tokens that publish models, and the API identities that query models.
| Profile item | Requirement (summarised) | Identity control that implements it |
|---|---|---|
| PO.5.1.R3 | Protect training pipelines, model registries and other components with least privilege | Dedicated workload identities per pipeline stage; registry write limited to the release pipeline identity |
| PO.5.1.R4 | Continuously monitor training activity in pipelines and model modifications in the registry | Audit logs on registry writes attributed to a specific identity, alerting on human or unexpected writers |
| PO.5.1.R6 | Continuously monitor each environment for plaintext secrets | Secret scanning on notebooks, training configs and container images; vaulted, short-lived credentials |
| PO.5.1.R1 | Monitor, track and limit resource usage and rates for model users during development | Per-identity quotas and rate limits on development endpoints |
| PO.5.1.C1 | Consider separating execution environments (isolation, segmentation, API-based access) | Separate cloud accounts or projects and identities per environment; no shared credentials across them |
| PS.1.1.R2, PS.1.1.R4 | Least privilege to minimise direct access to models; permit only indirect access to weights | Humans and applications reach weights through a serving or registry API, never raw storage keys |
| PS.1.1.C1, PS.1.1.C2 | Consider preventing all human access to weights; development only in approved environments | Break-glass role with approval and session recording; conditional access restricting where development identities work |
| PS.1.3.R3, PS.1.3.R4 | Least privilege on weights; encryption, hashes, signatures, multiparty authorisation and air gaps as risk warrants | KMS keys scoped to the serving identity; two-person approval for weight export |
NHI Mgmt Group's reading: the profile does not give specific guidance on API key management or on authenticating agents to models, so pair it with NHI controls for the credentials involved. See the AI Supply Chain and AI-BOM Guide for model and dataset provenance records.
Implementation steps
- Confirm your SSDF baseline: the profile assumes SSDF 1.1 practices are in place. Close basic gaps (protected repositories, build integrity, vulnerability response) first.
- Classify your role per product: model producer, system producer or acquirer. A system producer building on a hosted model will weight PW.4.4, RV.1.1 and PO.1.3 over PS.1.3.
- Load the profile table into your control register: import the tasks with their priorities, then re-rank to your risk and record the rationale for any change.
- Protect weights and data first (PS.1.2, PS.1.3): inventory where weights, checkpoints and datasets live; remove direct human and broad service access; enable integrity monitoring and access logging.
- Harden training environments (PO.5.1, PO.5.3): separate identities and accounts per environment, least-privilege pipeline roles, secret scanning, continuous monitoring with alert thresholds.
- Add data integrity checks (PW.3): poisoning and tampering analysis before training, provenance records where provenance is known, adversarial samples in test sets. See the AI Data Poisoning and Backdoors Guide.
- Gate acquired models (PW.4.4): verify signatures or hashes and provenance, scan for malicious content and vulnerabilities in an isolated environment, and record approval in the model registry.
- Extend testing and response (PW.8, RV): adversarial and red team testing, retest after retraining, input and output logging, and documented rollback criteria. See the AI Red Teaming Guide.
Rollout and rollback: pilot the profile on one model or AI product line for a quarter, measure the evidence it produces, then extend. If a control (for example blocking human access to weights) stalls research, grant a time-boxed, approved exception through a break-glass role rather than reopening standing access. Small versus large organisations: a small team consuming hosted models can focus on PW.4.4, RV.1.1, PO.5.1.R6 and supplier requirements under PO.1.3. A frontier or large model producer needs the full PS.1.3 set, including multiparty authorisation and isolated environments for weights.
Evidence and assessment
SP 800-218A has no certification scheme. Evidence is whatever demonstrates each task to an internal auditor, a customer or an assessor. Useful artefacts:
| Task area | Evidence |
|---|---|
| PS.1.2, PS.1.3 | Access lists for weight and dataset storage showing no direct human access; KMS key policies; integrity monitoring alerts and reviews |
| PO.5.1, PO.5.3 | Environment and identity architecture; least-privilege role definitions; secret scanning reports; monitoring dashboards with thresholds |
| PS.2.1 | Model signatures or hashes recorded in the registry; verification logs at load time |
| PW.3 | Data screening reports, provenance records (or documented unknowns, which the profile acknowledges), adversarial test set composition |
| PW.4.4 | Acquired model intake records: source, hash or signature check, scan results, approver |
| PW.8, RV.1, RV.2 | Red team and adversarial test reports, retest after retraining, input and output logging configuration, rollback criteria and drills |
Metrics: percentage of production models with signed weights verified at load; number of identities with direct read access to weights (target near zero); percentage of acquired models passing intake before use; time from detected model vulnerability to rollback or fix; percentage of training environments with continuous secret scanning.
Mapping to other frameworks
The profile itself cites the AI RMF, the OWASP Top 10 for LLM Applications and NIST AI 100-2 as informative references. The table below is NHI Mgmt Group's practical mapping for teams that track controls across frameworks.
| SP 800-218A area | Related framework | Reference |
|---|---|---|
| PW.3.1 data poisoning analysis | OWASP Top 10 for LLM Applications 2025; MITRE ATLAS | LLM04 Data and Model Poisoning; AML.T0020 Training Data Poisoning |
| PW.4.4 acquired models and components | OWASP LLM 2025; MITRE ATLAS | LLM03 Supply Chain; AML.T0010 AI Supply Chain Compromise; AML.T0115 Publish Poisoned AI Artifacts |
| PS.1.3 protect weights | MITRE ATLAS; NIST SP 800-53 Rev. 5 | AML.T0035 AI Artifact Collection; AC-6 Least Privilege; SI-7 Integrity |
| PS.2.1 model integrity | NIST SP 800-53 Rev. 5 | SR-4 Provenance; SI-7 |
| RV.1.1 input and output monitoring | OWASP LLM 2025 | LLM02 Sensitive Information Disclosure; LLM10 Unbounded Consumption |
| Whole profile | NIST CSF 2.0 | PR.PS-06 secure development practices; GV.SC-01 supply chain risk management programme |
| Risk framing | NIST AI RMF and AI 600-1 | Informative references in the profile; see the NIST AI 600-1 Generative AI Profile Guide |
Common pitfalls
- Using the profile alone: skipping the SSDF 1.1 base, so AI-specific controls sit on an unprotected build and repository estate.
- Treating weights like ordinary build artefacts: broad read access to object storage for data scientists and service accounts, contrary to PS.1.1 and PS.1.3.
- Secrets in notebooks and training configs: cloud keys and registry tokens in plaintext, which PO.5.1.R6 explicitly calls out.
- Trusting downloaded models: pulling a model from a public hub into a privileged environment without the PW.4.4 checks.
- Assuming the profile covers production: deployment and operation are out of scope, so runtime guardrails, agent permissions and inference monitoring need separate controls.
- Copying priorities without review: NIST presents them as a starting point; an acquirer and a frontier lab should not have the same ranking.
Practitioner checklist
- SSDF 1.1 baseline assessed and gaps tracked before adopting the profile.
- Organisational role (model producer, system producer, acquirer) recorded per AI product.
- Profile tasks imported into the control register with priorities reviewed and justified.
- Inventory of model weights, checkpoints and datasets with owners and storage locations.
- No standing direct human access to production weights; access via APIs and approved break-glass only.
- Pipeline and registry identities least-privileged, separated per environment, and monitored for unexpected writes.
- Secret scanning running across training environments, notebooks and images.
- Models signed or hashed at release and verified at load; acquired models pass intake checks.
- Data poisoning screening, provenance records and adversarial test sets in place for training data.
- Model rollback criteria defined and tested.
Standards and references
- NIST: SP 800-218A Secure Software Development Practices for Generative AI and Dual-Use Foundation Models
- NIST: SP 800-218A (PDF)
- NIST: SP 800-218 Secure Software Development Framework Version 1.1
- NIST: Secure Software Development Framework project
- OWASP: Top 10 for LLM Applications 2025
- MITRE ATLAS: Training Data Poisoning
- MITRE ATLAS: AI Supply Chain Compromise
Related NHI Mgmt Group resources: NIST Secure Software Development Framework (SSDF, SP 800-218) Guide · NIST AI RMF Guide · AI Supply Chain and AI-BOM Guide · Model Signing and OpenSSF Model Signing Guide. This guide summarises NIST SP 800-218A and is not legal advice.