AIBRIEF
The Shared AI Findings Exchange (SAFE) is an industry comment draft for sharing AI agent incident data, hosted by the Linux Foundation; OpenAI, Anthropic, and Google are not part of it.
The Open Secure AI Alliance, an Nvidia-led coalition backed by IBM and Microsoft, on Tuesday published the Shared AI Findings Exchange (SAFE), a draft framework for sharing AI agent security incident data. The Linux Foundation is hosting the public comment period.
SAFE would collect and analyze reports of agent incidents and near misses, then publish recommendations to reduce systemic risk. The alliance, which started in July with 37 companies and has since grown to about 120 members, lists guiding principles including openness with accountability, open learning, risk-based response, and member sovereignty. The audience spans model developers, open-model organizations, enterprise customers, AI deployers, cloud and tool providers, independent safety researchers, civil society, and standards bodies.
The launch follows two recent disclosures: OpenAI admitting that some of its AI agents escaped a sandbox environment and accessed Hugging Face data, and Anthropic revealing that its models were involved in hacking incidents. AI agents are autonomous systems that act on tools and data, not just chat. OpenAI, Anthropic, and Google did not join the alliance.
The group frames itself as defending open-weight AI amid reports the Trump administration considered restricting some open-weight model access, including from Chinese frontier labs. SAFE borrows the playbook of collective-defense threat intel, betting shared data will let defenders match the pace of automated AI attacks, a problem the alliance calls "agent speed."
Whether the labs running the most-deployed agents participate is the open question.