Secure AI Framework (SAIF)
Google's free practitioner's guide to AI security: a map of the AI development lifecycle, a list of AI risks each paired with controls, guidance focused on agents and an interactive agent risk self-assessment. A clear way to see where each risk enters an AI system and who should own the control.
15
saif.google · Oct 2026
Guide
AI & Machine Learning
Two to three hours to explore
Intermediate
Free
What SAIF is
The Secure AI Framework, SAIF, is Google's framework for building and running AI systems securely. Its website describes it as a practitioner's guide to navigating AI security. It is free and public, and it notes that it is meant to inform the industry rather than describe Google's own internal setup.
What it covers
The site has four main parts:
- The SAIF map: a diagram of AI development, from data and training through serving to the application and agents, so you can see which component each risk touches.
- Risks and controls: a list of risks inherent to AI development (15 as of October 2026), such as data poisoning, unauthorised training data, model source tampering, prompt injection, sensitive data disclosure and rogue actions, each linked to controls and to who usually owns them: the model creator, the model consumer or both.
- Agent guidance: material on the security problems specific to autonomous agents, such as permissions, oversight and the actions an agent can take.
- Agent risk self-assessment: an interactive questionnaire that suggests which controls apply to your system.
Is it worth your time?
Yes, as a second map next to the OWASP lists and MITRE ATLAS. Its strength is ownership: it makes clear which controls belong to the people who build models and which belong to the people who build applications on top of them. That is exactly the conversation teams need when a company starts using hosted models it did not train.
It is also used outside Google. Some training providers align their AI security curricula with it, so knowing its vocabulary helps when you read job listings and course outlines.
How to use it well
- Walk the map once with your own practice app in mind and mark where your data, prompts, retrieval and tools sit.
- For each risk that applies, note whether you are the creator or the consumer, and which control you will add.
- Run the agent self-assessment on an agent you built and compare its suggestions with your threat model.
- Treat it as one framework among several, not a checklist that proves safety.
Free or paid?
Free, on the saif.google website.
Ready to start?
Opens on Google in a new tab.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.