CAISI Works with OpenAI and Anthropic to Promote Secure AI Innovation

25 Sep 2025 · NIST – AI News (topic 2753736) US

A concrete US-UK government-industry AI security evaluation precedent is emerging - Australian AISI and DISR should assess how Australian arrangements compare.

Key points

Summary

NIST's Center for AI Standards and Innovation (CAISI) has conducted security evaluations of OpenAI and Anthropic systems under America's AI Action Plan, identifying vulnerabilities and supporting measurable security improvements. The evaluations were completed in partnership with the UK AI Security Institute, with both companies publishing blog posts describing concrete security improvements. CAISI and the UK AISI are continuing to collaborate on best practices in metrology, standards development, and advanced model capability assessment. This US-UK bilateral model of government-industry AI security testing is increasingly formalised and may shape international expectations for similar arrangements.

Implications for Australian agencies

Implications are AI-generated. Starting points, not advice.