Welcome to Super AI Safety
We are a home for clear thinking about the risks and responsibilities that come with increasingly capable artificial intelligence.
AI systems are moving from tools that answer questions to agents that act on our behalf: writing code, managing data, making decisions and operating in the world. As their capabilities grow, so does the importance of getting safety right. Super AI Safety exists to help people understand what is at stake and what can be done about it.
Here you will find plain-language explanations of the core challenges, including alignment, security, misuse, oversight and the governance questions that sit around them. We follow developments in research, regulation and industry practice, and we translate them into insights that leaders, practitioners and curious readers can use.
We avoid hype and alarmism, giving weight to what is known, what is uncertain and what still needs to be tested. Whether you are building AI, deploying it in your organisation, advising on its legal and ethical dimensions or simply trying to understand it, our aim is to give you reliable guidance.
Explore the resources, follow the conversation and join us in the work of making advanced AI safe, accountable and beneficial for everyone.
Book A Consultation
✓ Easy and hassle-free procedure
✓ Flawless business functions
-
Australia’s Digital Duty of Care
Australia’s proposed digital duty of care is one of the most important online safety reforms now moving through the policy pipeline. It is not just another content takedown scheme. It is an attempt to shift online safety law from a reactive model, where platforms respond after harm has already happened, to a systems-based model, where…
-
Which models achieve the highest bench mark scores with the lowest api cost?
The single most cost-effective frontier model on the market is Anthropic’s Claude Sonnet 5.5, while DeepSeek V4.1 Flash offers the best absolute value per dollar for near-frontier baseline performance. When balancing high benchmark intelligence with low API cost, the AI market splits into three distinct value tiers: true premium frontier systems optimized for efficiency, enterprise…
-
What is the rate of hallucination in frontier ai models?
There is no single hallucination rate for frontier AI models, as a model’s tendency to invent information depends entirely on the complexity of the task, the domain, and whether the system is “grounded” in source documents. [1, 2] Across industry benchmarks, the aggregate hallucination rate for frontier models (such as GPT-6, Claude 5/5.5, and Gemini…