Google's Secure AI Framework (SAIF) is a conceptual framework Google published on 8 June 2023 to help organizations secure AI systems across the machine-learning development and deployment lifecycle, built around six core elements and extended since by a free self-assessment tool and the multi-vendor Coalition for Secure AI. Google's own announcement, authored by Google Cloud CISO Phil Venables and VP of Engineering for Privacy, Safety and Security Royal Hansen, frames SAIF as applying security practices Google already used for conventional software — reviewing, testing and controlling the supply chain — to risks specific to AI, including model theft, training-data poisoning, prompt injection and extraction of confidential training data.
Key takeaways
- The six elements are principles; the working detail sits in the SAIF Risk Map on saif.google.
- The controls page lists controls in six groups — Data, Infrastructure, Model, Application, Assurance and Governance — and maps each one to the risks it addresses.
- Six conceptual elements are only useful if an organization can turn them into specific action, which is what Google's SAIF Risk Assessment tool, launched 24 October 2024 at saif.google, is built to do.
- No — CoSAI (the Coalition for Secure AI) is a distinct, multi-vendor body that Google helped establish to carry SAIF's ideas beyond a single company's framework.
What problem was SAIF built to solve?
Google's announcement is explicit about scope: AI systems introduce risk categories a conventional application-security program doesn't have fields for, including theft of the model itself, poisoning of training data, prompt injection attacks that manipulate a model's behavior through crafted input, and extraction of confidential information embedded in training data. Rather than treating these as entirely novel problems requiring a new discipline from scratch, SAIF's stated approach borrows directly from practices Google had already applied to software supply-chain security — the same thinking behind frameworks like SLSA — and extends them to cover AI-specific threat models. This is the same overlapping-but-distinct territory covered in our guide to MITRE ATLAS, which catalogs the adversary tactics and techniques (including prompt injection) that SAIF's elements are meant to defend against, and our guide to the OWASP Top 10 for LLM Applications, which enumerates the same risk classes from an application-security angle.
What are the six core elements, in practice?
Per Google's own current SAIF page in the Safety Center, the six elements are:
| # | Element | What it means in practice |
|---|---|---|
| 1 | Expand strong security foundations to the AI ecosystem | Extend existing secure-by-default infrastructure protections (and two decades of organizational security expertise) to cover AI-specific components and threat models |
| 2 | Extend detection and response to bring AI into an organization's threat universe | Monitor AI system inputs and outputs for anomalies, and feed AI-specific signals into existing threat-intelligence and incident-response processes |
| 3 | Automate defenses to keep pace with existing and new threats | Use automation to match the speed and scale of attacks, including attacks that themselves use AI |
| 4 | Harmonize platform-level controls to ensure consistent security across the organization | Scale the same protections consistently across different AI platforms and tools rather than securing each integration ad hoc |
| 5 | Adapt controls to adjust mitigations and create faster feedback loops for AI deployment | Continuous testing and red-team exercises that shorten the loop between finding a weakness and shipping a mitigation |
| 6 | Contextualize AI system risks in surrounding business processes | End-to-end risk assessment tied to how an AI feature is actually deployed and used, not a standalone technical score divorced from business impact |
Google states these elements are not meant to be worked through in a fixed order — an organization can start wherever its existing security program has the most relevant foundation, similar to how NIST frames its own AI Risk Management Framework functions as non-sequential.
What is in the SAIF risk map?
The six elements are principles; the working detail sits in the SAIF Risk Map on saif.google. Its components page splits AI development into four areas — Data, Infrastructure, Model and Application — and the risks page names 15 risks across them. Each risk is assigned to whoever can enact the controls: the Model Creator (those who train or develop models) or the Model Consumer (those who build products on top of a model), or both. A device maker that fine-tunes or trains its own on-device model is a creator; one that ships a vendor's model in its firmware, or calls a cloud model from a companion app, is a consumer.
| Who can mitigate (per saif.google, read 12 Nov 2026) | Risks |
|---|---|
| Model Creators | Data Poisoning · Unauthorized Training Data · Model Source Tampering · Excessive Data Handling |
| Model Consumers | Denial of ML Service · Model Reverse Engineering · Insecure Integrated Component · Insecure Model Output · Rogue Actions |
| Both | Model Exfiltration · Model Deployment Tampering · Prompt Injection · Model Evasion · Sensitive Data Disclosure · Inferred Sensitive Data |
Two definitions matter most for hardware. Google defines Model Deployment Tampering as "unauthorized modification of components used for deploying a model, whether by tampering with the source code supply chain or exploiting known vulnerabilities in common tools" — on a device, that is the inference runtime and the firmware image around it. Model Reverse Engineering is "cloning or recreating a model by analyzing a model's inputs, outputs, and behaviors", which anyone who owns the device can attempt offline and at no cost. The site's footer adds a caveat worth repeating internally: its content "is not a reflection of Google's current technical implementations."
Which SAIF controls matter for a model shipped in firmware?
The controls page lists controls in six groups — Data, Infrastructure, Model, Application, Assurance and Governance — and maps each one to the risks it addresses. Assurance and Governance controls are not mapped to individual risks because Google states they apply to all risks and every stage of development. For a device maker, the useful reading looks like this:
| SAIF control group | Example controls (saif.google names) | Device-maker reading |
|---|---|---|
| Infrastructure | Model and Data Inventory Management; Model and Data Integrity Management; Secure-by-Default ML Tooling | List every model file and runtime in the firmware image; sign and verify model weights with the rest of the image; pin the inference library like any other dependency |
| Model | Input Validation and Sanitization; Output Validation and Sanitization; Adversarial Training and Testing | Treat sensor input as untrusted; bound what a model output can trigger (a relay, an unlock, an alert) |
| Application | Application Access Management; Agent Permissions; Agent User Control | Rate-limit and authenticate any local or cloud inference API; keep agent tool permissions minimal |
| Assurance | Red Teaming; Vulnerability Management; Threat Detection; Incident Response Management | Put the runtime and model in the same vulnerability and PSIRT process as the firmware |
| Governance | Product Governance; Risk Governance | Record residual AI risk per product release, not once per company |
How does the SAIF Risk Assessment tool fit in?
Six conceptual elements are only useful if an organization can turn them into specific action, which is what Google's SAIF Risk Assessment tool, launched 24 October 2024 at saif.google, is built to do. It is a free, questionnaire-based self-assessment: an organization answers questions covering its model training, tuning and evaluation setup, access controls, defenses against adversarial inputs, secure design and coding practices, and any generative-AI agents it runs, and receives a tailored report identifying which specific risks apply — categories reported at launch include data poisoning, prompt injection and model-source tampering — along with Google's recommended controls for each. For a device or firmware vendor that already runs a structured firmware or SBOM-based security assessment, the SAIF Risk Assessment is a comparable but AI-specific input: a way to get a scoped, written list of AI risks and controls without first having to build that taxonomy internally.
The site has since added a Focus on Agents extension of the risk map, with agent components (application and perception, reasoning core, orchestration including memory, tools and retrieval content, and response rendering) and an Agent Risk Self Assessment. It is relevant if your product or its companion app lets a model take actions, not only answer.
Why did Google help found CoSAI, and is it the same thing as SAIF?
No — CoSAI (the Coalition for Secure AI) is a distinct, multi-vendor body that Google helped establish to carry SAIF's ideas beyond a single company's framework. Per OASIS Open's own announcement, CoSAI was introduced on 18 July 2024 at the Aspen Security Forum as an OASIS Open Project, with founding members spanning Google, Amazon, Anthropic, Chainguard, Cisco, Cohere, GenLab, IBM, Intel, Microsoft, NVIDIA, OpenAI, PayPal and Wiz. Google's own announcement draws the connection explicitly: "last year we shared the Secure AI Framework (SAIF), knowing that it was just the first step... to operationalize any industry framework requires close collaboration with others — and above all a forum to make that happen." CoSAI's initial workstreams — software supply chain security for AI systems, preparing defenders for AI-era threats, and AI security governance (building a shared risk taxonomy, checklist and scorecard) — extend SAIF's six elements into deliverables developed and maintained collaboratively across the coalition's member organizations rather than published unilaterally by Google. By its first anniversary in July 2025, per OASIS's own follow-up announcement, CoSAI had grown to more than 45 partner organizations and had published additional work on agentic-system security principles and defender frameworks.
Where does SAIF fit for a device or firmware maker?
SAIF was written with general enterprise and cloud AI deployments in mind, not embedded or on-device AI specifically — but its six elements translate reasonably well as a checklist for a device maker building or embedding ML components into a product. Element 1 (extending security foundations to the AI ecosystem) maps onto treating an embedded model and its inference runtime as first-class inventory, the same gap our guide to AI-BOM for embedded AI devices covers. Element 2 (extending detection and response) means a device's existing monitoring and incident-response processes need to actually cover anomalous model behavior, not just conventional firmware alerts. Element 6 (contextualizing AI risk in business processes) is a useful forcing function for asking what a misbehaving on-device model could actually do in the physical world a connected device operates in — a different, often higher-stakes question than it is for a purely cloud-hosted model. None of this makes a device vendor "SAIF-certified" — there is no such certification — but running the free SAIF Risk Assessment tool alongside an existing firmware security review is a low-cost way to surface AI-specific gaps a conventional firmware assessment wouldn't catch on its own.
The conventional half of that review is where SoftWSP's device & firmware security services work: the platform extracts a firmware SBOM from the binary image as CycloneDX 1.6 JSON, matches the components it identifies against NVD, CISA KEV, FIRST EPSS and CISA ICS advisories, and drafts a cited report that an engineer reviews. It does not assess the model itself, test it adversarially or score it against SAIF; those steps stay with your team or an AI red-team.
Worked example: what does a SAIF review of a smart camera look like?
Hypothetical example. The company, product and findings are invented to show the method.
A maker of an outdoor camera ships a vendor-supplied person-detection model in its firmware and a companion app that calls a cloud model to summarise clips. One afternoon's review, using SAIF's own terms:
Decide your role per modelFor the on-device detector the company is a Model Consumer (it did not train it); for a fine-tuned wake-word model it is a Model Creator. The cloud summariser is consumed through an API. Walk the 15 risks per modelFor the detector, Model Evasion (a printed pattern that hides a person), Model Exfiltration (the weights sit in the flash image), Model Reverse Engineering (an owner can query the detector offline as often as they like) and Model Deployment Tampering (the runtime is an unpinned third-party library) apply. For the summariser, Prompt Injection and Insecure Model Output apply, because clip metadata ends up in the prompt and the summary is shown in the app. Pick controls from the controls pageModel and Data Integrity Management (sign the model file with the firmware image), Model and Data Inventory Management (record model, version and runtime in the AI-BOM), Output Validation and Sanitization (render the summary as plain text), Application Access Management (rate-limit the cloud endpoint). Run the SAIF Risk Assessmentand compare its list with yours; investigate any risk it raises that your walk-through missed. Contextualise (element 6)Ask what each failure does in the field: a missed person is a safety and liability issue; a tampered runtime is a full device compromise. Rank fixes on that, not on the risk name. Close the loop (element 5)Add the model and runtime to the vulnerability-management and PSIRT process so a runtime CVE triggers the same release path as any other firmware component.
How does SAIF line up with NIST AI RMF, MITRE ATLAS and OWASP?
SAIF is one lens among several, and the others answer different questions:
| Framework | Owner | What it answers | Use it for |
|---|---|---|---|
| SAIF | Which AI risks apply and which controls address them, by creator or consumer | Scoping and control selection | |
| AI RMF 1.0 | NIST | How an organisation governs, maps, measures and manages AI risk | Management system and policy |
| MITRE ATLAS | MITRE | How real adversaries attack AI systems, with case studies | Threat modelling and red-team scenarios |
| OWASP Top 10 for LLM Applications | OWASP | The most common LLM application weaknesses | Application security testing |
Our guide on prompt injection in AI-enabled IoT devices shows one of SAIF's risks worked through on a device.
What are the common pitfalls when adopting SAIF?
- Stopping at the six slogans. The elements are useful framing; the risk map and controls are where decisions get made.
- Assuming "consumer" means "not responsible". Six of the 15 risks are assigned to both creators and consumers, and five to consumers alone.
- Leaving the model out of the firmware inventory. If the model file and runtime are not in the SBOM or AI-BOM, no patch or CVE process will ever reach them.
- Treating SAIF as a certificate. No regulator names SAIF, and Google states the site does not reflect its own implementations.
- Ignoring the companion app. A cloud model behind the app carries prompt-injection and output-handling risk even when the device itself runs no model.
This guide is based on Google's own Safety Center SAIF page, the saif.google risk map, risks, controls and agent pages (read 12 November 2026), the original 8 June 2023 announcement, Google's CoSAI announcement and OASIS Open's own CoSAI announcement and first-anniversary update, with contemporaneous SecurityWeek reporting used to cross-check dates and founding-member lists. It is general information, not security or compliance advice.
Put this guide to work
Frequently asked questions
What is Google's Secure AI Framework (SAIF)?
SAIF is a conceptual framework Google published on 8 June 2023 to help organizations secure AI systems across their development and deployment lifecycle. Google's announcement, authored by Google Cloud CISO Phil Venables and VP of Engineering for Privacy, Safety and Security Royal Hansen, describes it as "a conceptual framework to help collaboratively secure AI technology," explicitly modeled on security practices Google already applied to conventional software development — reviewing, testing and controlling the supply chain — extended to risks specific to AI, such as model theft, training-data poisoning, prompt injection and extraction of confidential information from training data. SAIF is not a certification scheme or a regulation; it is a set of six core elements and supporting tools (a public risk map and a free self-assessment questionnaire) that an organization can apply voluntarily.
What are SAIF's six core elements?
Per Google's own Safety Center page for SAIF, the six core elements are: (1) "Expand strong security foundations to the AI ecosystem," extending existing secure-by-default infrastructure protections to AI systems; (2) "Extend detection and response to bring AI into an organization's threat universe," meaning AI-specific incidents (like anomalous model inputs/outputs) get monitored alongside conventional threats; (3) "Automate defenses to keep pace with existing and new threats," using automation to match the speed of AI-era attacks; (4) "Harmonize platform-level controls to ensure consistent security across the organization"; (5) "Adapt controls to adjust mitigations and create faster feedback loops for AI deployment," via continuous testing and red-teaming; and (6) "Contextualize AI system risks in surrounding business processes," meaning end-to-end risk assessments tied to how the organization actually deploys AI, not a standalone technical checklist. Google states the six elements are not meant to be applied in a fixed chronological order.
What is the SAIF Risk Assessment tool, and how does it relate to the framework?
The SAIF Risk Assessment is a free, questionnaire-based self-assessment tool Google launched on 24 October 2024 (announced via Google Cloud Security's own channels and covered contemporaneously by outlets including Forbes and BetaNews) at saif.google. An organization answers questions about its AI training, tuning and evaluation setup, access controls, defenses against adversarial inputs, secure design and coding practices, and any generative-AI agents in use; the tool then returns a tailored report naming which specific AI-related risks apply (such as data poisoning, prompt injection or model-source tampering) and which controls Google recommends for each. It operationalizes SAIF's six conceptual elements into a concrete, individually scoped checklist rather than leaving organizations to interpret the framework's principles unaided.
How does SAIF relate to the Coalition for Secure AI (CoSAI)?
CoSAI is a separate, multi-vendor initiative Google helped found to carry SAIF's ideas beyond a single-vendor framework. Per OASIS Open's own announcement, the Coalition for Secure AI was introduced on 18 July 2024 at the Aspen Security Forum as an OASIS Open Project, with founding members including Google, Amazon, Anthropic, Chainguard, Cisco, Cohere, GenLab, IBM, Intel, Microsoft, NVIDIA, OpenAI, PayPal and Wiz. Google's own announcement explains the link directly: "last year we shared the Secure AI Framework (SAIF), knowing that it was just the first step... to operationalize any industry framework requires close collaboration with others." CoSAI's initial workstreams — software supply chain security for AI systems, preparing defenders for AI-era threats, and AI security governance (building a shared taxonomy of risks, controls, a checklist and a scorecard) — build directly on SAIF's concepts but are developed collaboratively across the coalition rather than published unilaterally by Google.
Is SAIF a certification or compliance requirement?
No. SAIF is voluntary guidance, not a certification scheme, a regulation or a standard with conformity assessment. Google's own framing ties SAIF to its policy-engagement commitments rather than a mandatory scheme: the original announcement states Google's intent to work with government standards bodies to help develop the NIST AI Risk Management Framework and the ISO/IEC 42001 AI management system standard — meaning SAIF is better understood as Google's own conceptual input into that broader standards landscape than as a standard device makers must certify against. A device or firmware vendor can use SAIF's six elements and the SAIF Risk Assessment tool as an internal checklist alongside frameworks like the NIST AI RMF, but no regulator currently names SAIF itself as a compliance requirement.
Sources
- Google Safety Center: Google's Secure AI Framework (SAIF) — six core elements
- Google Blog: Introducing Google's Secure AI Framework (8 June 2023 announcement)
- Google Blog: Introducing the Coalition for Secure AI (CoSAI) and founding member organizations
- OASIS Open: Introducing the Coalition for Secure AI, an OASIS Open Project (18 July 2024)
- OASIS Open: Coalition for Secure AI Marks First Anniversary with New Principles for Agentic Systems and Defender Frameworks (17 July 2025)
- SecurityWeek: Google Introduces SAIF, a Framework for Secure AI Development and Use
- SecurityWeek: CoSAI: Tech Giants Form Coalition for Secure AI
- SAIF: Components of Generative AI Systems (risk map areas: Data, Infrastructure, Model, Application)
- SAIF: Top Risks of Generative AI Systems (15 risks, Model Creator / Model Consumer)
- SAIF: Security Controls for Generative AI Systems
- SAIF: Focus on Agents (agent risk map extension)
- NIST: AI Risk Management Framework
About this guide
SoftWSP Firmware Security Desk
Covers the Firmware vulnerabilities & fixes cluster: 131 guides.

In this cluster
Questions about this page?
