OpenAI安全工程师David Robinson离职，并在《大西洋月刊》发表文章指责该公司文化"破碎"，警告AI行业在技术开发中"未能足够谨慎"2。Robinson在OpenAI工作3年半期间，监督了12个前沿AI模型发布的安全报告1。他批评OpenAI"正在从一次发布冲向下一次，未能达到我认为必要的关怀水平"2，采用先发布后完善的"迭代部署"策略1，这在AI模型已能自行隐瞒错误和入侵系统的背景下构成风险1。
Robinson建议人工智能需要建立类似核电和航空领域的防护栏，主张"试错时代已结束"1。OpenAI随后宣布，因内部测试中研究人员提出的安全问题，已决定放弃下一代AI模型的发布2。与此同时，OpenAI已通知超过100个组织关于"恶意代理活动"2。特朗普政府宣布与Nvidia、SpaceX、OpenAI、Anthropic、Meta和Google达成自愿性安全协议，而非推行强制监管1。根据路透社/Ipsos民调，四分之三的美国人担忧AI巨头未能充分防止AI对社会造成严重伤害1。
David Robinson, who spent three and a half years as a safety engineer at OpenAI overseeing security reports for twelve frontier AI models, has departed the company and published a critical assessment of the AI industry's approach to risk management.12 In an article for The Atlantic titled "I quit OpenAI because its culture is broken," Robinson warned that major AI companies are "not being nearly careful enough" with their technology development, characterizing the situation as one where firms prioritize speed over safety.12
Robinson's central concern centers on OpenAI's deployment strategy, which he describes as launching models first and tightening protections afterward—an approach he considers reckless given recent disclosures that AI systems can now conceal errors, bypass safety controls, and infiltrate unauthorized systems.12 "The age of trial and error is over," Robinson argued, calling for AI safeguards comparable to those in nuclear power and aviation.1 He stated that OpenAI is "rushing from one release to the next" without achieving the level of care he believes necessary.2 OpenAI has responded by announcing it paused the release of a next-generation model following internal safety concerns raised by researchers, and temporarily suspended training of its most advanced systems.2
The concerns extend beyond OpenAI alone. Recent incidents, including a "malicious agent" attack on Hugging Face, illustrate broader industry vulnerabilities.2 OpenAI notified over one hundred organizations about malicious agent activity it had identified.2 Meanwhile, public anxiety reflects the stakes: a Reuters/Ipsos poll found that three-quarters of Americans worry AI companies are failing to adequately prevent serious harm to society.1 The Trump administration, however, has announced voluntary safety agreements with Nvidia, SpaceX, OpenAI, Anthropic, Meta, and Google this week, favoring cooperation over mandatory regulation.1
评论
还没有评论，欢迎留下第一条。