We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Sep 12·5 min read
Working with US CAISI and UK AISI to build more secure AI systems
TL;DR
AI systems lack standardized safety and security practices, creating risks when deployed at scale. OpenAI partnered with US CAISI (Center for AI Safety and Innovation) and UK AISI (AI Safety Institute) to develop shared frameworks and best practices for secure AI development.
✦ Why It Matters
Engineers can adopt these coordinated safety frameworks to systematically evaluate and reduce risks in AI systems before production deployment.
Key Takeaways
- Implement red-teaming practices from CAISI/AISI frameworks to identify vulnerabilities before deployment.
- Adopt shared safety evaluation methodologies across teams to standardize risk assessment.
- Establish vulnerability disclosure processes aligned with government institute guidelines.
- Use coordinated frameworks to benchmark AI system safety against international standards.
Full Summary
Governments and AI developers have struggled to align on safety standards as AI systems become more powerful and widely deployed. OpenAI worked with two government-backed safety institutes—the US CAISI and UK AISI—to create shared methodologies for identifying and mitigating AI risks.
The partnership focused on developing practical frameworks that could be applied across different AI systems and organizations, rather than theoretical guidelines. Through collaborative workshops and technical exchanges, the teams documented best practices for red-teaming (adversarial testing), vulnerability disclosure, and safety evaluation.
The resulting frameworks are now being adopted by both institutes to assess AI systems before deployment. This work demonstrates that international coordination on AI safety standards is feasible and can produce concrete, implementable guidance.