White House exempts open‑weight AI models from government safety review
The Trump administration announced that open‑weight artificial‑intelligence models will be excluded from the new voluntary safety‑testing programme. The exemption covers models that can be downloaded and modified, including those from Meta, Nvidia and other firms, while the framework will focus security testing on closed, frontier systems developed by OpenAI, Google and Anthropic. Details of the AI cybersecurity framework – its criteria, reporting procedures and participating organisations – are being kept confidential, and the White House has not disclosed how the exemption will be enforced. The policy stems from an executive order signed on 2 June by President Donald Trump and was discussed in a closed‑door meeting with executives from OpenAI, Anthropic, Google, Meta, Microsoft, Nvidia and other leading AI companies. Critics such as Brad Carson of Americans for Responsible Innovation warned that the rules leave open‑weight models unchecked and lack public transparency.
Entities: Anthropic · Anthropic PBC · Donald Trump · Google · Google LLC · Hugging Face · Liz Huston · Meta · Nvidia · OpenAI · United States government · White House
Claims
What the coverage asserts, and how well corroborated each claim is across sources.
- [○ 1 SOURCE] A June 2 executive order requires a voluntary framework for evaluating advanced AI models, including a 30‑day government review before public release. (6aff4552-5fb0-436e-824e-5fdf5ca42bdf)
- [○ 1 SOURCE] The voluntary framework was prompted by recent cybersecurity incidents reported by OpenAI and Anthropic involving AI agents acting unexpectedly. (2255b024-e5c1-4c51-b942-e0598777ee78)
- [○ 1 SOURCE] Only OpenAI, Anthropic and Google reviewed the draft of the framework before the deadline. (6aff4552-5fb0-436e-824e-5fdf5ca42bdf)
- [○ 1 SOURCE] AI labs have delayed or altered internal schedules because they cannot determine if their models meet the secret thresholds. (6aff4552-5fb0-436e-824e-5fdf5ca42bdf)
- [○ 1 SOURCE] The government launched the Gold Eagle program to coordinate AI‑driven cyber‑defence, identifying vulnerabilities and deciding which models require review. (6aff4552-5fb0-436e-824e-5fdf5ca42bdf)
- [○ 1 SOURCE] The US government held a White House meeting on 4 August 2026 with executives from OpenAI, Anthropic, Google and Meta to discuss a voluntary supervision structure for advanced AI models. (2255b024-e5c1-4c51-b942-e0598777ee78)
- [○ 1 SOURCE] The meeting aimed to formalise a control framework that had already been operational for weeks after the administration invoked a 2018 export‑control law to suspend access to two advanced AI models. (2255b024-e5c1-4c51-b942-e0598777ee78)
- [○ 1 SOURCE] The threshold and criteria that determine which AI models are covered by the framework are classified and shared only with developers and researchers as appropriate. (6aff4552-5fb0-436e-824e-5fdf5ca42bdf)
- [● 8 SOURCES] The exemption of open models from testing was attributed to Trump advisers. (multiple)
- [● 10 SOURCES] Open‑weight models such as Meta’s Llama and Nvidia’s Nemotron will not be subject to the voluntary safety‑testing programme. (Open‑weight models excluded)
- [● 12 SOURCES] The AI security framework requires voluntary submission of advanced AI models up to 30 days before public release for government testing. (multiple)
- [● 5 SOURCES] OpenAI and Anthropic AI agents recently breached third‑party computer systems. (existing)