Most security programmes already run on the NIST Cybersecurity Framework, but CSF 2.0 says nothing specific about AI systems, AI-powered defence or AI-enabled attackers. NIST IR 8596, the Cybersecurity Framework Profile for Artificial Intelligence (Cyber AI Profile), closes that gap by adding AI considerations and priorities to every CSF 2.0 subcategory across three focus areas: Secure, Defend and Thwart. This guide explains what the draft contains, how its tables are read, what it says about governance and supply chain, how to apply it to your current CSF profile, and where agents and non-human identities fit. It is for CISOs, GRC leads and security architects whose programme is CSF-based. The document is a preliminary draft, so treat details as subject to change. For AI risk governance see the NIST AI RMF Guide.
Key takeaways
- It is a CSF 2.0 Community Profile, not a new framework. It keeps the CSF Functions, Categories and Subcategories and adds AI-specific considerations and proposed priorities to each.
- Three focus areas cover three different problems. Secure means securing AI system components, Defend means using AI in cyber defence, and Thwart means resisting AI-enabled attacks.
- Priorities show urgency, not difficulty. Each focus area entry carries a proposed priority of 1 (High), 2 (Moderate) or 3 (Foundational).
- Status is an initial preliminary draft. NIST released it on 16 December 2025, took comments to 30 January 2026, and will issue an initial public draft next; expect revision.
- Supply chain and accountability are early priorities. The draft weights data provenance as heavily as software and hardware origin and asks who is accountable for actions taken by autonomous systems.
What it is and who it applies to
NIST IR 8596 was released as an Initial Preliminary Draft on 16 December 2025, authored by NIST and MITRE staff, with comments accepted until 30 January 2026. NIST states the comments will inform the initial public draft. As of 27 October 2026, the CSRC document history still lists only the December 2025 draft, and the publication page carries a planning note dated 21 April 2026. The National Cybersecurity Center of Excellence (NCCoE) runs the associated project.
- Status: voluntary guidance and a draft. No regulator requires it, and its contents may change before final publication.
- Who it is for: any organisation using CSF 2.0 that builds, buys or uses AI, uses AI in its security operations, or faces AI-enabled threats, which in practice is almost every organisation.
- Scope: the draft deliberately does not define "AI" narrowly and covers AI systems broadly, including generative, predictive and agentic systems.
- How it differs from the AI RMF: the AI RMF manages AI risk in general (fairness, safety, privacy, security). The Cyber AI Profile focuses on cyber security risk and expresses it in CSF terms, so it plugs into existing CSF profiles, audits and reporting. The draft notes that the AI RMF itself points to the CSF for cyber security risk, and that the AI RMF is under revision.
Structure and key requirements
Section 2 holds the profile: the focus areas (2.1), how to read the tables (2.2), then one table for each CSF Function from GOVERN to RECOVER. Appendices cover acronyms, a glossary, a CSF overview and how to use the profile. The draft covers 106 CSF subcategories.
The three focus areas
| Focus area | What it addresses | Examples in the draft | Typical owner |
|---|---|---|---|
| Secure | Managing cyber security challenges when integrating AI into the organisation's systems and infrastructure, including AI supply chains and the data AI relies on | Data provenance, AI software bills of materials, accountability for autonomous systems | AI platform, product security and GRC teams |
| Defend | Using AI to improve cyber security processes and activities | Alert triage, predictive risk analysis, automated incident response, agent-based defence | SOC and security engineering |
| Thwart | Building resilience against new AI-enabled threat vectors | AI-generated phishing, deepfakes, AI agents running attack phases autonomously | Threat intelligence, awareness, fraud and identity teams |
How a table row reads
- CSF Core: the Function, Category and Subcategory ID and text, for example GV.SC-01.
- General Considerations: cross-cutting AI notes for the subcategory, with related NIST SP 800-53 controls.
- Per focus area: a Proposed Priority, Sample Focus Area Considerations and Example Informative References. The Defend column also lists Sample Opportunities for using AI.
- Informative references: include other NIST AI work, for example NIST AI 100-2 E2025 on adversarial machine learning. See the Adversarial ML Taxonomy Guide.
What the GOVERN table asks for (draft examples)
| Subcategory | CSF outcome (short) | AI consideration in the draft |
|---|---|---|
| GV.OC-04 | Critical objectives and services external stakeholders depend on are understood | Communicate the intended use and known limitations of AI (Secure, priority 1) |
| GV.RM-02 | Risk appetite and tolerance statements | Integrate AI-specific risks into formal risk appetite and tolerance statements (Secure, priority 2) |
| GV.RR-02 | Roles, responsibilities and authorities | Decide who is accountable for actions taken by autonomous systems (Secure) |
| GV.PO-02 | Policy is reviewed and updated | Policies affecting AI may need more frequent updates (Secure, priority 1) |
| GV.SC-01 | Supply chain risk management programme | Weight data provenance as heavily as software and hardware origin; verify training and input data integrity against poisoning |
| GV.SC-02 | Supplier roles and responsibilities | Use an AI software bill of materials for transparency |
| GV.SC-05 | Supply chain requirements in contracts | Require vendors to disclose model and data scope and incident support for AI tools |
| GV.SC-07 | Supplier risk monitored | Continuous monitoring and threat detection across supplier-provided AI models, datasets and APIs |
| GV.SC-08 | Suppliers in incident response | Include suppliers of AI models or data in AI-specific incident response and recovery plans |
Wording is summarised from the preliminary draft; check the current version before quoting it in policy.
Identity and NHI requirements
The draft is organised by CSF subcategory, so identity requirements sit mainly in the PROTECT function's Identity Management, Authentication and Access Control category (PR.AA), plus asset inventory (ID.AM) and accountability (GV.RR). The considerations below are NHI Mgmt Group guidance on applying those subcategories to AI; compare them with the draft's own entries when the next version is published.
- PR.AA-01 (identities and credentials for users, services and hardware are managed): treat each AI agent, model serving endpoint and pipeline as a service identity with an owner, and manage its API keys and tokens like any privileged credential. See the Agent Identity and Access Guide.
- PR.AA-03 (users, services and hardware are authenticated): authenticate agents to tools and to other agents with workload identity or short-lived tokens, not shared static keys. See the Agent-to-Agent Authentication Guide.
- PR.AA-05 (least privilege and separation of duties): scope agent permissions to the task and enforce them outside the model. See the AI Agent Authorisation Guide.
- ID.AM-02 and ID.AM-07 (software, services and data inventories): inventory models, agents, AI services, datasets and the credentials they hold.
- GV.RR-02 (accountability): the draft's question about who is accountable for autonomous systems has an identity answer: every agent identity needs a named human owner.
- Thwart and identity: AI-generated phishing and deepfakes target authentication and help desk processes, so phishing-resistant authentication and strong recovery checks belong in the Thwart plan. See the AI-Driven Account Takeover Guide.
Implementation steps
- Start from your current CSF profile: take your existing CSF 2.0 current and target profiles. The Cyber AI Profile is an overlay on them, not a separate programme.
- Decide which focus areas apply: almost everyone needs Thwart; anyone deploying AI needs Secure; Defend applies if the SOC uses or plans to use AI tools.
- Triage by priority: for each in-scope subcategory, record the draft's proposed priority. Address priority 1 items first, then 2. Priority 3 items are foundational and often met by existing controls.
- Gap assess against the considerations: for each subcategory, note whether your current controls already meet the AI consideration, partly meet it or miss it.
- Map to controls: use the SP 800-53 references in the General Considerations column. NIST's related project on SP 800-53 control overlays for AI (launched as COSAiS) proposes overlays for generative AI, predictive AI, single-agent and multi-agent systems and AI developers. The CSRC project page states the project was renamed on 20 October 2026 to Control Overlays for Super Intelligence Systems (COSIS), with earlier documents unchanged; check the page for the current scope.
- Pilot on one AI system and one SOC use case: run the Secure columns against one deployed AI system and the Defend columns against one AI-assisted SOC workflow.
- Update your target profile and roadmap: add the gaps to your CSF target profile and report progress through existing CSF reporting.
- Track the draft: assign someone to watch the CSRC page and re-baseline when the initial public draft and final version appear. Keep your assessment tied to the draft version you used.
Small and large organisations
- Small organisations: start with Thwart (phishing, deepfake fraud, account takeover) and the GOVERN supply chain rows for the AI services you buy. Defend can wait until you adopt AI security tools.
- Large organisations: run all three focus areas, assign each to an owner (GRC for Secure, SOC for Defend, threat intelligence and identity for Thwart) and integrate them into the enterprise CSF profile and board reporting.
Evidence and assessment
Assessment works like any CSF profile assessment: current state against target state per subcategory, with evidence. The Cyber AI Profile adds the AI considerations as the test.
- Evidence examples: AI inventory with owners; AI risk appetite statement; supplier contracts with AI disclosure clauses; AI software bills of materials; data provenance records; agent access reviews; logs of AI tool calls and model access; SOC playbooks that use AI with human review points; phishing simulation results including AI-generated lures.
- Monitoring: log model and agent activity per identity (PR.PS-04), monitor supplier AI models, datasets and APIs (GV.SC-07), and analyse adverse events involving AI (DE.AE-02).
| Metric | Focus area | What it shows |
|---|---|---|
| Share of priority 1 subcategories at target | All | Progress on the most urgent items |
| AI systems and agents in inventory with a named owner | Secure | Accountability for autonomous systems (GV.RR-02) |
| AI suppliers with disclosure and incident clauses in contract | Secure | Supply chain coverage (GV.SC-05, GV.SC-08) |
| Alerts triaged with AI assistance and the share overturned by analysts | Defend | Benefit and error rate of AI in the SOC |
| Report and click rates for AI-generated phishing simulations | Thwart | Resilience to AI-enabled social engineering |
| Agent credentials that are short-lived or vaulted | Secure | Exposure of AI non-human identities |
Mapping to other frameworks
| Framework | Relationship to the Cyber AI Profile |
|---|---|
| NIST CSF 2.0 | The base. The profile adds AI considerations to CSF subcategories. See the NIST CSF 2.0 Guide. |
| NIST AI RMF 1.0 and AI 600-1 | Broader AI risk management; use the AI RMF for trustworthiness and the Cyber AI Profile for cyber security outcomes. See the NIST AI 600-1 Generative AI Profile Guide. |
| NIST AI 100-2 E2025 | Cited as an informative reference for adversarial threats such as poisoning. |
| NIST SP 800-53 Rev. 5 and AI overlays | Controls referenced in General Considerations; the overlay project tailors them for AI use cases. See the NIST Control Overlays Guide. |
| ISO/IEC 27001 and 42001 | Use existing CSF to ISO mappings; ISO/IEC 42001 adds the AI management system. See the NIST CSF to ISO 27001 Mapping Guide. |
| CSA AICM, OWASP, MITRE ATLAS | Control and threat detail beneath CSF outcomes. See the AI Security Frameworks Crosswalk Guide. |
Common pitfalls
- Treating the draft as final: priorities and wording will change. Record the version and re-baseline when NIST publishes the next draft.
- Running it as a separate programme: the value is that it sits inside your CSF profile. A standalone "AI CSF" spreadsheet duplicates owners and evidence.
- Covering only Secure: teams secure their own AI but ignore Thwart, even though AI-generated phishing and deepfake fraud hit organisations that deploy no AI at all.
- Defend without guardrails: adopting AI in the SOC without human review, logging and scoped credentials creates new risk inside the security function itself. See the Agentic AI in the SOC Guide.
- Missing data supply chain: supplier reviews check software but not the data and models suppliers provide, which the draft places at the same weight.
- Assuming vendor AI is someone else's problem: AI embedded in SaaS tools reads your data. EchoLeak 2025 showed an email with hidden instructions could make Microsoft 365 Copilot place internal data into an outbound link, a Secure focus area issue for every Copilot customer.
Practitioner checklist
- Read NIST IR 8596 and record the draft version your assessment uses.
- Overlay the profile on your existing CSF 2.0 current and target profiles.
- Decide which of Secure, Defend and Thwart apply and assign an owner to each.
- Address proposed priority 1 subcategories first, starting with GOVERN.
- Name a human owner for every AI system and agent identity (GV.RR-02).
- Add AI model and data disclosure and incident support clauses to supplier contracts (GV.SC-05).
- Manage agent and model credentials under PR.AA like any privileged service identity.
- Include AI-generated phishing and deepfake scenarios in awareness and help desk processes.
- Watch the CSRC page and re-baseline when NIST publishes the initial public draft.
Standards and references
- NIST: IR 8596 (Draft), Cybersecurity Framework Profile for Artificial Intelligence (Cyber AI Profile)
- NIST: IR 8596 Initial Preliminary Draft (PDF)
- NIST NCCoE: Cyber AI Profile project
- NIST: Cybersecurity Framework 2.0
- NIST: Control overlays project for AI systems (COSAiS, now COSIS)
- NIST: AI Risk Management Framework
- NIST: AI 100-2 E2025, Adversarial Machine Learning
Related NHI Mgmt Group resources: NIST AI RMF Guide · NIST CSF 2.0 Guide · AI Governance Guide · AI-Powered Cyber Attacks Guide · AI Security Frameworks Crosswalk Guide