谷歌、微软和xAI已同意在美国商务部机构CAISI的测试下，提前评估其前沿AI模型。测试的特殊之处在于，公司将提供降低或移除安全护栏的模型版本，以便评估其在协助网络入侵、恶意软件规划等高危任务上的原始能力与风险。此前，OpenAI和Anthropic已于2024年达成类似协议。此举背景是白宫正考虑建立针对主要AI模型的政府审查流程，审查重点是其网络能力——即发现和利用软件漏洞以改变现实安全风险的水平。政策转向的触发点是Anthropic的Mythos模型，该公司认为该模型在发现安全漏洞方面能力过强，广泛发布风险过高。
Google, Microsoft and xAI just agreed to let the U.S. government test early frontier AI models before the public can use them.
The testing will be run by CAISI, a Commerce Department group that checks what advanced models can do and where they may create security risk.
The unusual part is that the companies will share versions with reduced or removed guardrails, which lets testers see the model’s raw ability instead of only its polished public behavior.
Becasue, a national security test asks whether the model can help with cyber intrusion, malware planning, or other high-risk tasks when its filters are weakened.
CAISI has already completed more than 40 evaluations, including tests on models that have not been released.
OpenAI and Anthropic made similar agreements in 2024, so the new deal pulls Google, Microsoft and xAI into the same pre-release testing lane.
---
wsj .com/tech/ai/google-microsoft-and-xai-agree-to-share-early-ai-models-with-u-s-f95a88d1
Nytimes: The White House is considering a government review process for major AI models before public release. The proposed review would not necessarily block releases, but it could give the U.S. government early access to test advanced models before the public gets them. The likely focus is cyber-capability, meaning whether a model can help discover, exploit, or chain software vulnerabilities at a level that changes real-world security risk. This is a major policy reversal because the same administration had rolled back Biden-era reporting rules for powerful AI systems and argued that heavy regulation could slow U.S. progress against China. The trigger was Anthropic’s Mythos model, which the company says is unusually strong at finding software security flaws and too risky to release widely. A model like this can scan code, identify weak points, and suggest attack paths much faster than a normal human security team. That makes it useful for defense agencies, but it also raises the fear that a public release could help attackers find bugs in banks, government systems, cloud tools, and critical infrastructure. --- nytimes. com/2026/05/04/technology/trump-ai-models.html在 X 查看被引用的帖子
来源：Rohan Paul · x.com