Trump Signs Voluntary AI Safety Order as White House Races to Finalise Model Release Standards Before August Deadline
The Trump administration is in advanced talks with major artificial intelligence companies to codify voluntary standards for the release of frontier AI models, with an announcement expected imminently, the Financial Times reported on July 2. The discussions follow President Trump's June 2 executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security," which directs developers to voluntarily submit their most powerful models to federal agencies for up to 30 days of pre-release testing. The order tasks the NSA, CISA, and the National Institute of Standards and Technology with building a classified benchmarking process to assess when any given model crosses the threshold of being a "covered frontier model" — one whose cyber capabilities could pose national security risks.
According to Reuters, the proposed voluntary standards would set benchmarks for advanced models and establish timelines and access rules governing who can use them, both within the United States and abroad. The move arrives weeks after Washington's handling of Anthropic's Fable 5 model — which was pulled offline for nineteen days in June following a government export control order — threw into sharp relief how ad hoc federal intervention in AI releases can be. Anthropic itself has publicly called for the pre-release framework to be applied equally across all frontier model developers and codified in durable regulation, as reported by Build Fast with AI. The White House, Anthropic, and OpenAI did not immediately respond to requests for comment from Reuters.
The June 2 executive order is notably voluntary in character — a deliberate policy choice reflecting months of internal debate within the administration. As The Hill reported, the White House had previously considered a mandatory pre-deployment approval regime but reversed course over concerns that heavy-handed regulation would impede American competitiveness. OpenAI CEO Sam Altman, while complying with the administration's request to delay model releases, noted publicly that the current arrangement is not optimal, particularly objecting to the idea of government selecting who gets early access to new models. A White House official described the approach as one of collaboration with frontier labs rather than imposition.
The EU's AI Act adds a hard deadline that Washington's voluntary posture cannot defer. The Act's August 2, 2026 compliance obligations remain fully binding for any US company operating in European markets, regardless of what framework the White House ultimately finalises, according to Tech Jacks Solutions. The divergence creates an asymmetric compliance burden: multinationals must simultaneously satisfy Brussels' prescriptive mandates and navigate Washington's evolving, industry-negotiated norms. Analysts have noted that this regulatory split risks producing a de facto two-tier global AI ecosystem, with firms that rely on US government partnerships potentially excluded from certain European use cases.
The stakes of getting the framework wrong are high in both directions. AI-enabled cyberattacks are accelerating — the Five Eyes intelligence alliance warned in July that such threats are months away, according to Build Fast with AI — while over-restriction of frontier models risks ceding ground to less safety-conscious developers. Anthropic's experience with Fable 5 has already illustrated that blunt export controls can prove counterproductive: the Commerce Department lifted its ban on June 30 after Anthropic's own testing demonstrated that rival models from OpenAI and China's Moonshot AI could replicate the same exploit that triggered the original restriction. The coming weeks of negotiation between the White House and the labs will likely define the architecture of AI oversight for years to come, and the August deadlines on both sides of the Atlantic mean there is no room for further delay.