Strengthening our safeguards through collaboration with US CAISI and UK AISI (www.anthropic.com)

Anthropic has deepened its collaboration with the US Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute (AISI) to enhance the security of its AI models through rigorous third-party testing. This ongoing partnership gives government experts unprecedented access to Anthropic’s systems—including pre-deployment versions and various safeguard configurations—enabling them to identify and stress-test vulnerabilities in critical defense mechanisms like the Constitutional Classifiers used to detect and prevent jailbreak attempts. The collaboration uncovered significant weaknesses, such as prompt injection attacks, sophisticated universal jailbreaks, cipher-based obfuscation, and automated adversarial attack refinements, which prompted fundamental architectural improvements rather than simple patches. This work is significant for the AI/ML community as it demonstrates the power of public-private partnerships combining government expertise in national security and threat modeling with industry innovation in AI safety. The feedback from CAISI and AISI has not only fortified Anthropic’s immediate safeguard measures but also strengthened its broader risk assessment and mitigation strategies. By providing transparent access to documentation, safeguard scores, and iterative testing processes, Anthropic facilitated deeper insights into complex attack vectors that automated or external bug bounty programs might miss. The collaboration highlights the importance of sustained, transparent engagement to effectively address evolving threats in increasingly powerful AI systems, setting a precedent for future security initiatives across the industry.