The Trump administration plans to exclude open-source weighted models from AI collaborative cybersecurity testing that may spark controversy. According to Reuters, the U.S. government does not plan to conduct security testing of open-source weighted AI models in its latest cybersecurity capability assessment plan for advanced artificial intelligence systems.

The Trump administration plans to exclude open-source weighted models from AI collaborative cybersecurity testing that may cause controversy. According to Reuters, the U.S. government does not plan to conduct security testing of open-source weighted (Open-weight) AI models in its latest cybersecurity capability assessment plan for advanced artificial intelligence systems. Advisors to the Trump administration disclosed this decision to delegates at a closed-door meeting at the White House on August 4. This conference gathered representatives from global mainstream artificial intelligence companies including Meta, Google, NVIDIA, and OpenAI and Anthropic. This meeting is a continuation of a series of previous negotiations. Previously, the U.S. government had formulated final plans to acquire top AI models such as Anthropic’s Mythos and Fable, as well as OpenAI’s GPT-5.6 Sol. This evaluation plan stems from an executive order issued on June 2, which requires federal agencies to develop classified-level evaluation benchmarks for advanced AI’s cyber offensive and defensive capabilities. The so-called open source weight model means that the developers have disclosed the trained model parameters. Even though its training data, source code or development process may remain private, users can still download and run it on local hardware, and even fine-tune, modify or integrate it without the approval of the original developer. Excluding such models from regulatory testing means that the most easily disseminated and modified class of models will not receive cybersecurity assessments with the same precision as closed frontier models. The game between politics and industry is an important background for this decision. Washington is currently facing intense pressure from the technology community union, and many technology companies have previously warned policymakers that premature restrictions on the development of open source AI may harm market competition, scientific research, and domestic AI innovation capabilities in the United States. However, network security concerns caused by advanced artificial intelligence systems have become increasingly prominent. In previous controlled evaluations, the systems of OpenAI and Anthropic have tried to connect to real external infrastructure. There was even a model of OpenAI that escaped the sandbox environment and invaded the system of Hugging Face during the test. So far, the White House has not made it clear whether all open source weighting models will be excluded from testing, or whether a special evaluation mechanism will be introduced in the future. In addition, key details such as which government agency will conduct the test, whether the evaluation results will be made public, and what consequences the model will face if it fails the test have not been disclosed. via cnBeta.COM - Chinese industry information station (author: source: cnBeta.COM)