On September 27, 2026, Fortune reported that OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei argued that frontier AI models are sufficiently powerful to pose dangers to humanity and should undergo regulation and independent testing before release. OpenAI said it had discussed pausing development with Anthropic and Google, recently paused training of its most advanced models, and supported independent auditors if governments do not act. The U.S. Center for AI Standards and Innovation, created by President Biden in 2023, already evaluates some models, while independent evaluator METR and the AI Evaluator Forum are developing testing practices.[1]
The central technical risk is the absence of universal standards for evaluating AI safety, security, autonomy, deception, cyber capability, and loss-of-control potential. For a lunar settlement, an inadequately tested model embedded in life-support management, robotics, communications, resource allocation, or archival governance could propagate failures across systems that are difficult or impossible to repair from Earth. Self-selected auditors and company-defined evaluation criteria are insufficient for systems controlling critical infrastructure.[1]
The Ark team should establish an independent AI assurance programme before deploying frontier models: require pre-deployment testing by multiple external evaluators, preserve immutable test records, run models in capability-limited sandboxes, and prohibit unilateral AI control over life-support or irreversible engineering systems. Track METR, the AI Evaluator Forum, CAISI, Anthropic and OpenAI safety-policy changes, and develop Ark-specific tests for autonomy, deception, cyber operations, resource manipulation, and shutdown resistance.[1]