英伟达CEO黄仁勋近日连续发起AI安全倡议。他发布了由32家公司联署的公开信，主张AI安全防御需要依靠开源工具和模型[1]。随后，他宣布成立"开源安全AI联盟"(Open Secure AI Alliance)，该联盟获得37家创始成员的加入[1]。英伟达还开源了NOOA框架(NVIDIA Labs Object-Oriented Agent)，在SWE-bench Verified基准测试中达到82.2%的准确率，CyberGym L1解决率为86.8%，同时token消耗仅为竞品的一半[1]。
在这波倡议中，主要的大型AI公司态度各异。OpenAI和Google均在黄仁勋的公开信上署名，但两家公司都未选择加入新成立的联盟[1]。Anthropic的立场最为突出——该公司既未在公开信上签署，也未加入联盟[1]。Anthropic CEO Dario Amodei发表声明指出，公司不认为"开放能让防御者更安全"，反而主张通过芯片封锁、杜绝蒸馏和加强安全测试等手段来加强AI安全[1]。
NVIDIA CEO Jensen Huang has released an open letter and announced the formation of the Open Secure AI Alliance to advance AI safety through open-source tools and models [1]. The initiative has attracted 37 founding members, though responses from major AI companies vary significantly [1].
The open letter, co-signed by 32 companies, endorses the use of open-source models as a cornerstone of AI security defense [1]. NVIDIA has open-sourced the NOOA framework (NVIDIA Labs Object-Oriented Agent), achieving an 82.2% success rate on SWE-bench Verified and an 86.8% resolution rate on CyberGym L1, while consuming only half the tokens of competing solutions [1].
The response from leading AI companies has been mixed [1]. OpenAI and Google signed the public letter supporting the initiative but did not join the alliance [1]. Anthropic, however, declined to participate in either the letter or the alliance [1]. Anthropic CEO Dario Amodei released a statement opposing the open-source approach, arguing that "openness does not make defenders safer" [1]. Instead, Anthropic advocates for chip restrictions, prevention of model distillation, and strengthened safety testing [1].
