NIST SP 800-218A, "Secure Software Development Practices for Generative AI and Dual-Use Foundation Models: An SSDF Community Profile," is NIST's companion resource adding AI-model-specific practices and tasks on top of its Secure Software Development Framework (SP 800-218). Published July 2024 (DOI 10.6028/NIST.SP.800-218A) and authored by a team spanning NIST's Computer Security and Applied Cybersecurity Divisions, CISA, and Scarfone Cybersecurity, it was written to satisfy a specific legal mandate. Its own Section 1 states: Section 4.1(a) of Executive Order 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," tasked NIST with "developing a companion resource to the Secure Software Development Framework to incorporate secure development practices for generative AI and for dual-use foundation models" — and, in the document's own words, "This document is that companion resource."
Key takeaways
- It is a Profile.
- Yes, as a NIST publication.
- Many unchanged SSDF tasks gain AI-specific recommendations.
- Read "code" as "firmware image plus model file".
Is it a replacement for SP 800-218 or a Profile?
It is a Profile. The document says it does not stand alone: "This Profile supplements what SSDF version 1.1 already includes... and should not be used without SP 800-218." An SSDF Community Profile, in NIST's terminology, is "a baseline of SSDF practices and tasks that have been enhanced to address a particular use case" — here, AI model development. The Profile walks through every practice and task of SP 800-218 and tags each row of its Table 1 in one of three ways: unchanged from SSDF 1.1 (the majority, where a "No additions to SSDF 1.1" note points back to the base framework), "Modified from SSDF 1.1" (the task itself was adjusted), or "Not part of SSDF 1.1" (a new task). Every task also carries one of three priority tags: High ("critically important for AI model development security compared to other tasks"), Medium ("directly supports AI model development security"), or Low ("beneficial... but is generally not more important than most other tasks"). NIST frames these as "a starting point for organizations to assign their own priorities," not a fixed ranking.
Is SP 800-218A still current after EO 14110 was revoked?
Yes, as a NIST publication. The Executive Order that commissioned it is gone, but the CSRC landing page still lists SP 800-218A as final, and the SSDF project page names it as the NIST-provided Community Profile. The dates, from the Federal Register and CSRC:
| Date | Event |
|---|---|
| 30 Oct 2023 | EO 14110 signed; Section 4.1(a) tasks NIST with an SSDF companion resource for generative AI |
| Jul 2024 | SP 800-218A published as final, built on SSDF 1.1 |
| 20 Jan 2025 | EO 14148 revokes EO 14110 |
| 6 Jun 2025 | EO 14306 directs NIST to publish "a preliminary update to the SSDF" by 1 December 2025 and a final version within 120 days of it |
| 17 Dec 2025 | SP 800-218 Rev. 1 initial public draft (SSDF version 1.2) published; comments closed 30 January 2026 |
On 13 November 2026, CSRC lists that initial public draft as the latest SSDF 1.2 text. The Profile's task IDs refer to SSDF 1.1. When SSDF 1.2 is final, check whether a task ID you cite has moved. The Profile is voluntary guidance, so losing the Executive Order changes its origin story, not its content.
Who is it written for, and what is out of scope?
The Profile names three primary audiences: AI model producers (organisations developing their own generative AI or dual-use foundation models), AI system producers (organisations building software that leverages such a model), and AI system acquirers (organisations acquiring a product or service built on one or more AI systems). A device maker that fine-tunes a vendor's generative model and ships it in firmware is a model producer for the fine-tuning and a system producer for the product.
Its Scope section draws a hard line: "This Profile's scope is AI model development, which includes data sourcing for, designing, training, fine-tuning, and evaluating AI models, as well as incorporating and integrating AI models into other software. Consistent with SSDF version 1.1 and EO 14110, practices for the deployment and operation of AI systems with AI models are out of scope." Runtime risks in a deployed AI system or agent — prompt injection, excessive agency, insecure plugin design — sit outside this document; see our MCP server security and OWASP Top 10 for LLM Applications guides for that layer. The Profile also covers "only... cybersecurity risk management," pointing to NIST's separate AI Risk Management Framework for privacy, bias and other AI risk types.
What's actually new about protecting weights and training data?
Table 1 of the full text carries six tasks tagged "Not part of SSDF 1.1" and two tagged "Modified from SSDF 1.1":
| Task | Status | What it requires |
|---|---|---|
| PO.5.3 | Not part of SSDF 1.1 — High | "Continuously monitor software execution performance and behavior in software development environments to identify potential suspicious activity and other issues," including every environment that hosts model APIs, weights, configuration parameters or training datasets. |
| PS.1.2 | Not part of SSDF 1.1 — High | "Protect all training, testing, fine-tuning, and aligning data from unauthorized access and modification." |
| PS.1.3 | Not part of SSDF 1.1 — High | "Protect all model weights and configuration parameter data from unauthorized access and modification," with "risk-proportionate cybersecurity practices around model weights, such as encryption, cryptographic hashes, digital signatures, multi-party authorization, and air-gapped environments." |
| PW.3.1–PW.3.3 | New practice PW.3, three tasks | "Confirm the Integrity of Training, Testing, Fine-Tuning, and Aligning Data Before Use." PW.3.1: "Analyze data for signs of data poisoning, bias, homogeneity, and tampering"; PW.3.2 tracks data provenance where known; PW.3.3 includes adversarial samples in training and testing data. |
| PO.1.3 | Modified — Medium | Requirements communicated to third-party software suppliers; the AI recommendation adds AI model development security to them. |
| PS.3.2 | Modified — Medium | Provenance data for every component of each release, now naming SBOM and SLSA; recommends tracking the provenance of an AI model "and its components and derivatives, including the training libraries, frameworks, and pipelines." |
PS.1.1 is unchanged as a task but its recommendation extends code storage protection to "AI models, model weights, pipelines, reward models, and any other AI model elements." The Profile's introduction gives the reason: AI development "blurs traditional boundaries between system code and system data," training sets "may be acquired from unknown, untrusted sources," and model weights "can be susceptible to malicious tampering."
What do the existing tasks add for AI models?
Many unchanged SSDF tasks gain AI-specific recommendations. For device teams the most useful are:
- PS.2.1 (Medium): "Generate and provide cryptographic hashes or digital signatures for an AI model and its components, artifacts, and documentation," and "Provide digital signatures for AI model changes."
- PW.4.4 (High): "Verify the integrity, provenance, and security of an existing AI model or any other acquired AI components," training datasets included.
- PW.6.1 (Low): "Consider using secure model serialization mechanisms that reduce or eliminate vectors for the introduction of malicious content."
- PW.7.2 (High): "Scan all AI models for malware, vulnerabilities, backdoors, and other security issues."
- PW.8.2 (High): test all AI models for vulnerabilities and "Retest AI models when they are retrained or new data sources are added."
- RV.1.3 (Medium): "Include AI model vulnerabilities in organization vulnerability disclosure and remediation policies," and tell users about model limitations and how to report problems.
How does the Profile define a dual-use foundation model?
Its glossary reproduces the Executive Order's definition verbatim: an AI model "trained on broad data," that "generally uses self-supervision," "contains at least tens of billions of parameters," is "applicable across a wide range of contexts," and "exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety" — for example by "substantially lowering the barrier of entry for non-experts" to build CBRN weapons, "enabling powerful offensive cyber operations through automated vulnerability discovery and exploitation," or "permitting the evasion of human control or oversight through means of deception or obfuscation." The definition adds that "models meet this definition even if they are provided to end users with technical safeguards that attempt to prevent users from taking advantage of the relevant unsafe capabilities." The small generative models that fit on a device are far below that scale; for them, the generative AI half of the title is what brings their development into scope.
How does a device team apply it to a model shipped in firmware?
Read "code" as "firmware image plus model file". For a generative model shipped on the device, the tasks map onto steps most firmware teams already run:
| Profile task | Firmware-team action | Evidence to keep |
|---|---|---|
| PS.1.3, PS.2.1 | Sign the model file with the release key, verify it in the secure-boot or update chain, keep weights in a protected partition | Signing records, verification test |
| PS.3.2 | List the model, its runtime and conversion toolchain in the release SBOM or AI-BOM | SBOM per release |
| PW.4.4 | Record the source, licence and hash of any pre-trained model before fine-tuning | Acquisition record |
| PW.6.1 | Use a format without executable deserialization for shipped models | Format decision record |
| PW.3.1–3.3 | Screen fine-tuning data; keep provenance; add adversarial samples | Data review log |
| PW.7.2, PW.8.2 | Scan and test the model; retest after retraining | Test reports per model version |
| RV.1.3 | Add model issues to the disclosure policy and security.txt scope | Published policy |
For the SBOM side, see AI-BOM for embedded AI devices; for signing, OpenSSF model signing; for provenance levels, SLSA.
Worked example: what does this look like for a device with an on-device language model?
Synthetic example: the company, model and numbers are invented for illustration.
A maker of industrial HMI panels fine-tunes a small open-weight language model on its own manuals so operators can ask questions on the panel, offline. The quantized model ships inside the firmware.
Acquire (PW.4.4)The team records the upstream model's repository, release tag, licence and SHA-256 before use. Protect the data (PS.1.2, PW.3.1, PW.3.2)The 12,000 question-and-answer pairs built from the manuals sit in an access-controlled repository; a reviewer samples 2% for wrong or injected answers and logs the result; each batch's source document is recorded. Train and evaluate (PW.3.3, PW.8.2)Prompt-injection strings and unsafe-instruction requests are added to the test set; results are recorded per model version and rerun after every retraining. Package (PS.1.3, PS.2.1, PW.6.1)The weights are exported to a weights-only format with no executable deserialization, hashed and signed with the firmware release key; the bootloader rejects an unsigned model. Inventory (PS.3.2)The release SBOM lists the model file, the inference runtime and its version, and the conversion tool. Respond (RV.1.3)The disclosure policy names the assistant as in scope and the manual tells operators the model's limits and how to report a problem.
The prompt-injection behaviour of the shipped panel is a deployment question; the Profile covers how the model was built, not how it is run.
Where does it draw on other NIST and OWASP work?
The Profile cross-references entries to NIST's AI Risk Management Framework and NIST's adversarial machine learning taxonomy for most PW and RV tasks, and maps individual tasks to entries in the OWASP Top 10 for LLM Applications (version 1.1). Task RV.2.2 ("Plan and implement risk responses for vulnerabilities") carries a note that AI model risk responses "should consider the time and expenses that may be associated with rebuilding them," a cost that does not apply to patching conventional source code. NIST's Generative AI Profile (AI 600-1) covers the wider risk side.
What are the common pitfalls?
- Treating the model as data, not code. PS.1.1 puts models and weights under the same storage protection as source code.
- Signing the firmware but not the model. If the model loads from a separate partition, PS.2.1 and PS.1.3 apply to it on its own.
- Leaving the model out of the SBOM. PS.3.2 asks for provenance of "all components of each software release."
- Loading pickled models. PW.6.1 points to serialization formats that cannot carry executable content.
- Testing once. PW.8.2 asks for retesting after every retraining or new data source.
- Citing the Profile for runtime controls. Deployment and operation are out of scope.
Where should a team start?
Read it alongside SP 800-218, not instead of itThe Profile only makes sense layered onto the four base practice groups covered in our NIST SSDF guide. Do the High-priority new tasks first:PS.1.2, PS.1.3, PO.5.3 and PW.3.1. Respect the scope boundaryPair it with runtime guidance for deployed models. Use the priority tags as a starting pointNIST frames High/Medium/Low as "a starting point for organizations to assign their own priorities." Track SSDF 1.2Map your evidence to task IDs so a renumbering is a table update, not a rewrite.
SoftWSP covers one slice of this. Its firmware analysis extracts the software components it identifies in a firmware image into a CycloneDX 1.6 JSON SBOM (firmware SBOM), matches them against NVD, CISA KEV, FIRST EPSS and CISA ICS advisories, and drafts a cited report for an engineer to review. That supports the component-inventory part of PS.3.2 and vulnerability monitoring under RV.1. It does not inspect model weights, scan models for backdoors, review training data or test model behaviour; those Profile tasks stay with your team. See Device & Firmware Security.
This guide summarises NIST SP 800-218A's own text (abstract, Introduction, Scope, Table 1 and Appendix A glossary, read from the published PDF), its companion SP 800-218 and the Federal Register records cited above; it is not a substitute for the full publication and does not constitute compliance or legal advice.
Put this guide to work
Frequently asked questions
What is NIST SP 800-218A, and why was it written?
SP 800-218A, "Secure Software Development Practices for Generative AI and Dual-Use Foundation Models: An SSDF Community Profile," was published by NIST in July 2024 (DOI 10.6028/NIST.SP.800-218A), authored by Harold Booth, Murugiah Souppaya and Apostol Vassilev (NIST Computer Security Division), Michael Ogata (NIST Applied Cybersecurity Division), Martin Stanley (CISA) and Karen Scarfone (Scarfone Cybersecurity). Its own Section 1 states the mandate directly: Section 4.1(a) of Executive Order (EO) 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," tasked NIST with "developing a companion resource to the Secure Software Development Framework to incorporate secure development practices for generative AI and for dual-use foundation models." The document's own words: "This document is that companion resource."
Does SP 800-218A replace SP 800-218, or does it stand alone?
It stands alone only in combination with the base framework — the document is explicit that it "should not be used without SP 800-218." Its own Purpose section states: "This Profile supplements what SSDF version 1.1 already includes. The Profile is intended to be used in conjunction with NIST Special Publication (SP) 800-218... Readers should also utilize the implementation examples and informative references defined in SP 800-218 for additional information." See our [NIST SSDF (SP 800-218) guide](/guides/nist-ssdf-sp-800-218/) for the base framework's four practice groups — Prepare the Organization (PO), Protect the Software (PS), Produce Well-Secured Software (PW) and Respond to Vulnerabilities (RV) — that this Profile annotates rather than replaces.
Who is SP 800-218A written for, and what is explicitly out of scope?
The document names three primary audiences in its own Audience section: "AI model producers" (organisations developing their own generative AI and dual-use foundation models, "as defined in EO 14110"); "AI system producers" (organisations building software that leverages such a model); and "AI system acquirers" (organisations acquiring a product or service that uses one or more AI systems). Its Scope section is equally explicit about the boundary: "This Profile's scope is AI model development, which includes data sourcing for, designing, training, fine-tuning, and evaluating AI models, as well as incorporating and integrating AI models into other software. Consistent with SSDF version 1.1 and EO 14110, practices for the deployment and operation of AI systems with AI models are out of scope." Runtime AI-system security controls therefore sit outside this document's remit — this guide's [MCP server security guide](/guides/mcp-server-security/) and [OWASP Top 10 for LLM Applications guide](/guides/owasp-top-10-for-llm-applications/) cover deployed-system risk instead.
What AI-specific practices and tasks did the Profile add that are not in SSDF 1.1?
The Profile tags every row in its Table 1 as either unchanged, "Modified from SSDF 1.1," or "Not part of SSDF 1.1." Two tasks under practice PS.1 (Protect All Forms of Code From Unauthorized Access and Modification) carry that "Not part of SSDF 1.1" tag and a High priority: **PS.1.2**, "Protect all training, testing, fine-tuning, and aligning data from unauthorized access and modification," and **PS.1.3**, "Protect all model weights and configuration parameter data from unauthorized access and modification." PS.1.3's own recommendations include specifying "risk-proportionate cybersecurity practices around model weights, such as encryption, cryptographic hashes, digital signatures, multi-party authorization, and air-gapped environments." A whole new practice, **PW.3**, "Confirm the Integrity of Training, Testing, Fine-Tuning, and Aligning Data Before Use," is also new to the Profile; its first task, PW.3.1, reads: "Analyze data for signs of data poisoning, bias, homogeneity, and tampering before using it for AI model training, testing, fine-tuning, or aligning."
How exactly does EO 14110 define a "dual-use foundation model"?
The Profile's own glossary (Appendix A) reproduces the EO 14110 definition verbatim, citing it as source [1]: a dual-use foundation model is "An AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters," such as by "substantially lowering the barrier of entry for non-experts to design, synthesize, acquire, or use" CBRN weapons, "enabling powerful offensive cyber operations through automated vulnerability discovery and exploitation," or "permitting the evasion of human control or oversight through means of deception or obfuscation." The same entry adds: "Models meet this definition even if they are provided to end users with technical safeguards that attempt to prevent users from taking advantage of the relevant unsafe capabilities."
Sources
- NIST SP 800-218A: Secure Software Development Practices for Generative AI and Dual-Use Foundation Models — An SSDF Community Profile (full text PDF)
- NIST CSRC: SP 800-218A — publication landing page (abstract, keywords, DOI)
- NIST CSRC: SP 800-218, Secure Software Development Framework (SSDF) Version 1.1 — the base framework this Profile supplements
- NIST: Secure Software Development Framework (SSDF) project page
- NIST news: "Secure Software Development Practices for Generative AI and Dual-Use Foundation Models" — publication announcement, July 2024
- The White House: Executive Order 14110, Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (Section 4.1(a) mandate)
- Federal Register: Executive Order 14148 of 20 January 2025, Initial Rescissions of Harmful Executive Orders and Actions (revokes EO 14110)
- Federal Register: Executive Order 14306 of 6 June 2025, Sustaining Select Efforts to Strengthen the Nation's Cybersecurity (Sec. 2, SSDF update tasks)
- NIST CSRC: SP 800-218 Rev. 1 initial public draft, SSDF Version 1.2 (17 December 2025)
About this guide
SoftWSP Firmware Security Desk
Covers the SBOM cluster: 31 guides.

In this cluster
Questions about this page?
