AI Safety / Security
Build the future. Protect what matters.
AI can expand what people learn, create and achieve. Making that progress worthwhile also means asking who could be harmed, how a system might fail and what would help us keep control. At Saturdays.AI, learning to build goes hand in hand with learning to question.
Start with the foundations
The Introduction to AI Safety course explores advanced AI risks, malicious use, loss of control, systemic risks, evaluations, alignment, oversight and mitigations. It connects these ideas with real cases and possible future scenarios.
The course is taught in Spanish. Its current content and access details are available on the Saturdays.AI learning platform.
Explore the introductory course
Browse all courses on cursos.saturdays.ai
Two perspectives. One responsibility.
AI Safety asks how we can reduce harmful behaviour and unintended consequences, evaluate capabilities and limitations, and maintain meaningful human oversight.
AI Security also considers deliberate attacks: attempts to manipulate a system, expose sensitive information or misuse its tools and permissions. Protecting an AI application involves its data, software and operating environment as well as the model.
The two perspectives overlap. A useful starting question is: what could go wrong in this particular use, who would be affected, and what evidence would show that a safeguard works?
Bring these questions to your next project
- Define what the system should do and the actions it should never take on its own.
- Test realistic failures and misuse in an authorised environment, using data you are allowed to use.
- Limit access to data and tools, and decide where a person must review or approve an action.
- Document the limits, monitor what happens and be ready to pause or change the system.
These are starting points for learning. A successful prototype or a course certificate does not establish that a system is safe for every context.
Keep exploring
The NIST framework offers a structured way to consider risks throughout an AI system's lifecycle. OWASP's guidance introduces practical security risks and mitigations for generative AI applications. These are independent resources for further learning.
Have a question or an idea to explore with others? Contact the community or discover community projects.