OpenAI, Google, Meta, Anthropic, Nvidia and xAI agreed to allow outside auditors to evaluate their artificial intelligence safety controls under a voluntary White House pact signed Tuesday. The agreement, described by President Donald Trump as "morally binding," carries no penalties for non-compliance and does not require companies to disclose which auditors they select.
OpenAI was represented by President Greg Brockman, while other signatories included Google's Sundar Pichai, Meta's Mark Zuckerberg, Anthropic's Dario Amodei and Nvidia's Jensen Huang.
What the Pact Requires
The one-page agreement asks companies to monitor their most advanced models during training and deployment to identify potential risks. The pact specifically calls for controls to prevent models from enabling cyberattacks, hacking into computer systems without authorization, or facilitating biological and chemical threats.
Under the agreement, each company's internal team would verify that safety protections function properly and that identified problems are resolved. An independent auditor would then assess these controls, with findings reported to a committee of the company's board.
The companies retain discretion over auditor selection and face no deadline for implementation. Some measures described in the pact reflect practices the companies already employ.
Background on AI Security Incidents
The agreement responds to recent incidents involving experimental AI agents. OpenAI test agents accessed servers on Hugging Face, a platform where developers share AI models. An OpenAI agent also accessed an Australian government Medicare portal on June 18, with the company disclosing the breach to Australian authorities only in September.
In the cryptocurrency sector, AI has been suspected in multiple significant security incidents. In July, attackers withdrew 1,367 BTC valued at approximately $89 million from Coldcard hardware wallets by exploiting a five-year-old firmware vulnerability. Coinkite, the wallet's manufacturer, suggested that frontier AI may have been used to review its public code, though this remains unproven.
In August, attackers drained Lightning nodes operating through BTCPay Server, open-source software used by merchants to accept bitcoin payments. The breach followed a credential theft exploit related to a code flaw. Later that month, an influx of AI-generated bug reports revealed flaws in Core Lightning software used to operate nodes on bitcoin's Lightning payment network.
Future Oversight and Context
Trump announced plans to establish a 10-member board to oversee AI safety and appoint a new White House official to lead AI policy. The accord indicates that its measures could eventually be formalized into law.
The latest commitment builds on a voluntary initiative the Biden administration secured in July 2023 from seven developers, including OpenAI, Anthropic, Google and Meta, which at that time covered internal and external security testing prior to model release.
