Verification Mechanisms
Tools and protocols that prove whether AI systems meet their safety claims across training, evaluation and deployment, promoting greater diffusion, international coordination and shared best practice.
Research
Practical research for AI's biggest coordination challenges.
Our Approach
Our research focuses on mitigating large-scale risks from frontier AI that require international coordination. We take a collaborative approach, bridging between stakeholders, regions and technical & policy solutions.
Our work aims to be solutions-oriented and practical, emphasizing technical proofs of concepts over theory, such that we can catalyse action by the wider ecosystem.
Featured Research
Recent publications and working papers from SASH and our collaborators.
Detecting Offensive Cyber Agents
AI agents can now orchestrate cyberattacks, changing the speed and scale of cyber threats. This report proposes detection-in-depth as a strategic framework for defenders and policymakers.
Evaluating DeepSeek v4 Pro For Frontier Risk
We carry out an independent safety evaluation of DSv4 Pro across CBRN, cyber, harmful manipulation, and loss-of-control. Our findings show that while capability is near-frontier; safeguards do not hold under trivial attack.
Building the trust layer for AI agents
As AI agents begin to take action in the real world, the services and people they interact with need to know who they are, who instructed them, what they are allowed to do and what should happen if something goes wrong. We propose a design logic to direct the development of agent IDs.
Research Focus
Frontier Safety
Safety evaluations that can check model safeguards for frontier risks such as Loss of Control and Harmful Manipulation in a diverse range of deployment contexts, before models move into public and commercial use.
Agent Governance
Solutions that can identify, supervise and constrain autonomous agents across regions, actors and use cases, enabling users to integrate them confidently into high-stakes and dynamic use cases.
Interested in doing the work?
We are looking for partners, researchers and builders with a solutions mindset and strong bias towards action.
Contact UsSee Open Roles(hiring globally!)