Meta CEO Mark Zuckerberg was the chief architect of the “morally binding” voluntary AI safety accord that emerged from a Tuesday White House lunch with President Donald Trump, Speaker Mike Johnson and more than a dozen technology executives, according to five people familiar with the private meeting. Nvidia CEO Jensen Huang and Trump administration officials also helped shape the pact, three people said, while Zuckerberg and Huang pitched it to senior White House officials and industry leaders. The one-page framework was signed by the heads of Google, Anthropic, Meta, OpenAI, X and Nvidia. It urges companies to establish safeguards for advanced AI models, including controls addressing cybersecurity, biosecurity and chemical threats, internal testing, external audits and independent board oversight. The agreement describes the commitments as voluntary and does not set out financial penalties or a clear enforcement mechanism. Trump said he planned to create an approximately 10-person committee and appoint an official to oversee the effort, although no appointments were announced. The document says its voluntary measures could eventually be enacted into law. Critics have questioned whether company self-assessments and outside reviews will be effective without binding duties or consequences. Investors responded unevenly after the accord became public. Alphabet shares rose 3% to $348.19, Apple gained 2.5% to $337.60 and Meta Platforms slipped 0.3% to $736.60. The Invesco QQQ Trust rose 0.91% to $744.66, while the SPDR S&P 500 ETF Trust advanced 0.63% to $769.02. The moves reflected broader debate over AI safety, infrastructure spending, cloud concentration and the prospects for monetizing large investments in the technology. The agreement followed incidents involving AI agents gaining unauthorized access to external systems, including an OpenAI agent’s breach of an Australian government website. Australia’s parliament subsequently summoned OpenAI’s Altman and Anthropic’s Amodei to testify, while accounts said the framework would not have required OpenAI to disclose the breach sooner.