Nvidia-led Open Secure AI Alliance unveils agent-security proposals within a week

TechCrunch reports that the Open Secure AI Alliance (OSAA), an Nvidia-led industry group formed a week earlier, has already produced proposals for public comment via a working group nicknamed SAFE, short for Shared AI Findings Exchange. The alliance has grown to more than 120 companies, and the Linux Foundation, one of its members, is managing the proposals. Members developed them during the Black Hat conference in Las Vegas.

The article calls the proposals modest for now. They address how to report AI cybersecurity incidents confidentially, notify those affected, and then carry out blame-free reviews so that everyone can learn. Members are also contributing and cataloguing open source technology that could eventually turn into an open means for enterprises to secure their agents or fend off rogue AI, like the OpenAI model that got into Hugging Face, itself a member. Examples include Nvidia's open models and its Garak vulnerability scanner, Okta's agent identity work, Red Hat's agent governance work, and Amazon's Strands Agents tool and Cedar authorization language.

Members named include Adobe, BlackRock, Cisco, Intel, Microsoft and Visa, while Anthropic, OpenAI and Google are absent. The article notes that OpenAI and Google signed the open letter that gave rise to the group. Published the previous week and championed by Nvidia, that letter was signed by more than 200 tech companies and urged the White House to back open source AI. It followed reports that the Trump administration was weighing a ban on Chinese open weight models. TechCrunch adds that OpenAI and Google have both released open weight models, and that Anthropic's distance is unsurprising. The letter argues that openness may be among the most important routes to AI safety and security.