How Google DeepMind, Microsoft & xAI Voluntarily Align with NIST AI Safety

Three of the most powerful AI labs in the world have agreed to let the US government stress-test their models before public release. NIST‘s Center for AI Standards and Innovation (CAISI) announced this week that Google DeepMindMicrosoft and xAI have signed voluntary pre-deployment safety testing agreements covering national security risk assessment. The backdrop is a regulatory vacuum: President Trump rescinded Executive Order 14110 in January 2025, removing the prior federal mandate for frontier AI safety provisions and leaving voluntary commitments as the main lever of government oversight.

That policy gap matters more than the headline. Without a binding federal framework, what these labs choose to do, and what CAISI actually tests, carries weight that would otherwise be distributed across enforceable rules. What follows is an account of how frontier labs are implementing that commitment, from policy navigation through to post-deployment accountability.

Phase 1: Navigating the Evolving Policy Landscape

The first challenge for any frontier AI lab is keeping track of what the rules actually are, given they keep changing.

1. Deconstructing Current U.S. AI Policy

With EO 14110 rescinded, the direct federal mandate for AI safety provisions has been replaced by EO 14179, “Removing Barriers to American Leadership in Artificial Intelligence,” which reorients policy toward sustaining US AI dominance while addressing potential model bias. CAISI continues to develop non-regulatory standards and guidelines, including the NIST AI Risk Management Framework (AI RMF), a voluntary guide covering four core functions: Govern, Map, Measure and Manage. Labs aligning their internal risk processes with the AI RMF gain a credible baseline for both domestic engagement and international regulatory conversations.

2. Engaging with Global Regulatory Frameworks

US policy shifts do not insulate labs from international obligations. The EU AI Act established the first comprehensive legal framework for AI, categorising systems by risk level and imposing stringent transparency and safety requirements on general-purpose models that pose systemic risks. For labs with global deployments, compliance with the EU framework often sets the floor for internal safety standards worldwide. Engagement with OECD principles and participation in international standards forums are, in practice, as operationally relevant as domestic regulatory filings.

Phase 2: Implementing Proactive Pre-Deployment Safety Measures

Before an advanced model ships, frontier labs are investing heavily in adversarial testing and documentation designed to catch problems that standard benchmarks miss.

1. Establishing Advanced Red-Teaming Protocols

Red-teaming means deliberately attacking your own system before an adversary does. Structured teams attempt to elicit harmful outputs, circumvent safety controls, or identify failure modes that developers did not anticipate. Though EO 14110 formally defined AI red-teaming and directed NIST to develop guidelines for it, the practice predates that order and has survived its rescission. Internal red teams challenge model assumptions from inside the lab; external red-teaming brings in independent researchers to stress-test mitigations against novel attack vectors. OpenAI has detailed its external red-teaming process for models including GPT-4 and DALL-E 3, and the approach is now standard practice across the frontier labs covered by the CAISI agreement.

2. Developing Model Transparency and Documentation

Comprehensive documentation, covering system architecture, decision-making criteria, validation outcomes and bias mitigation steps, is not just a compliance exercise. It is increasingly what government partners use to assess a model’s safety posture before deployment.

3. Integrating Secure Development Lifecycle Practices

Security has to be designed in from the start, not bolted on after. That means threat modelling specific to AI systems, identifying attack vectors like prompt injection, data poisoning and model extraction, then implementing mitigations before the model reaches a user. Google DeepMind’s Frontier Safety Framework takes this further, proactively identifying and mitigating severe risks as the lab’s understanding of model capabilities evolves. Protecting model weights from compromise is part of the same picture: the underlying infrastructure is as much a security surface as the model itself.

Phase 3: Ensuring Continuous Post-Deployment Assurance

Signing a pre-deployment testing agreement with CAISI is the beginning of the process, not the end. Real-world deployment surfaces risks that controlled testing does not.

1. Sustained Monitoring and Incident Response

Once a model is live, labs run continuous monitoring to detect performance drift, identify emergent misuse and catch anomalous outputs before they compound. Human review operates alongside automated tools. Incident response plans need to be functional before an incident occurs, including procedures for rapid patching, system decommissioning if required and thorough post-incident analysis. The CAISI partnership, with its national security focus, raises the stakes on response speed considerably.

2. Independent Audits and External Verification

Internal safety testing is necessary but not sufficient. The voluntary CAISI arrangement is itself a form of external verification: a government entity assessing models for national security risks before public launch, rather than relying solely on the lab’s own safety disclosures. Third-party assessments of safety claims, participation in public safety challenges and structured collaboration with academic researchers serve the same function, providing an independent check on whether internal measures are actually working.

3. Deploying Explainable AI for Accountability

Explainable AI (XAI) addresses a specific problem: advanced models can produce outputs that are correct without being interpretable. Techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) produce human-readable explanations of how a model reached a given output. In regulated sectors, that interpretability is not optional. A credit decision, a content moderation call, a hiring recommendation, each requires a defensible audit trail. Labs integrating XAI early in development build that trail into the system architecture rather than reconstructing it after the fact, which matters considerably when a regulator or a government safety partner asks to see the reasoning.

Phase 4: Contributing to Foundational AI Safety Research

Compliance with current requirements is the floor. The harder problem is building systems that remain safe as capabilities scale.

1. Investing in Interpretability and Alignment Research

Interpretability research aims to make a model’s internal processes legible, so engineers can identify where and why a system fails. Alignment research tackles the harder question: ensuring that increasingly autonomous models act in ways consistent with human intentions. Google DeepMind’s published work in this area includes safety fine-tuning and post-training harmlessness instruction, while also acknowledging the persistent challenge of jailbreaking, where users find ways to circumvent safety training after the fact. Neither problem is solved. The CAISI testing agreements are partly valuable because they create external pressure to keep working on both.

2. Collaborating on Open Standards and Best Practices

No single lab has a monopoly on safety insights. The Frontier Model Forum, co-founded by Google DeepMind, coordinates cross-industry work on safe frontier AI development. The Partnership on AI brings together academic institutions, civil society organisations and company labs on shared challenges. Joint red-teaming protocol design, shared benchmarks for measuring model behaviour under adversarial conditions and threat intelligence exchange all raise the collective safety baseline. This kind of cooperation sits in legal and competitive territory that requires careful navigation, but the alternative, every lab independently discovering the same failure modes, is slower and riskier.

3. Fostering a Culture of Responsible AI Development

Internal governance structures matter alongside external partnerships. Google DeepMind operates a Responsibility and Safety Council and an AGI Safety Council, both of which review research and projects against the lab’s stated AI principles. Training programmes, clear governance role assignments and documented accountability at each development stage are how safety commitments become operational practice rather than public relations. The CAISI partnership is most credible when it reflects something already happening internally, not something labs are adopting to satisfy an external obligation.

The voluntary CAISI agreements with Google DeepMind, Microsoft and xAI represent a calculated bet by those labs: that proactive engagement with government safety testing is preferable to waiting for mandatory frameworks to arrive on less favourable terms. The contrast with Anthropic is instructive. The lab’s reported dispute with the Pentagon over military use of its Claude model, said to have produced a legal challenge against a “supply chain risk” designation by March 2026, illustrates how sharply commercial interests and national security requirements can diverge. The labs that signed with CAISI are making a different calculation: that transparency now, through red-teaming, model documentation, XAI integration and foundational safety research, reduces the risk of a more adversarial regulatory encounter later. Whether that calculation proves correct depends on what CAISI’s testing actually surfaces. For more on the interplay between AI development and government oversight, see our coverage of the UK Frontier AI Taskforce report. For more coverage of AI policy and regulation, visit our AI Policy & Regulation section.