US AI Safety Institute Pacts with OpenAI and Anthropic Set Up Frontier Model Audit Standards

The U.S. Artificial Intelligence Safety Institute (AISI), housed within the National Institute of Standards and Technology (NIST), finalized formal Memorandums of Understanding with OpenAI and Anthropic. Under these agreements, the institute receives access to major upcoming foundation models from both labs prior to and immediately following public deployment. The collaborative framework establishes technical research partnerships focused on assessing advanced capabilities, uncovering emergent safety risks, and developing standardized risk mitigation feedback loops. This development transitions frontier LLM evaluation from self-reported proprietary safety cards to institutionalized, independent government evaluation. Enterprise platform teams integrating LLM APIs often operate without full visibility into the raw attack surfaces, jailbreak vulnerabilities, and dangerous capability thresholds of closed-source frontier models. Having a centralized institute conduct pre-release red-teaming provides an authoritative baseline that helps enterprise compliance, legal, and security officers evaluate downstream exposure. It also pressures providers to resolve systemic model misalignment before releasing APIs to enterprise customers. The initiative reflects the broader institutionalization of AI governance following executive and international safety summits, such as bilateral safety pacts between the U.S. and UK AI Safety Institutes. Historically, cloud and infrastructure engineering matured through standardized testing and external verification protocols—similar to NIST frameworks for cryptographic validation and cybersecurity maturity. By extending this paradigm to large language models, regulatory bodies and AI vendors are laying the groundwork for verifiable model governance akin to traditional enterprise software compliance. For DevOps and ML engineering teams, this signals the onset of formalized compliance checklists for LLM adoption. Practitioners should anticipate new evaluation rubrics filtering down into cloud AI marketplaces and enterprise service level agreements. In the near term, teams building LLM-backed workflows should not rely solely on provider safety promises. Instead, align internal red-teaming frameworks with emerging NIST evaluation benchmarks, monitor official AISI test findings when upgrading model versions, and implement robust application-layer defenses—such as guardrails and input-output validation—to manage risks that upstream testing may not fully catch.