Nvidia, Cisco Among 120+ Groups Proposing SAFE AI Agent Incident Reporting

More than 120 organizations, including Nvidia, Cisco and CrowdStrike, have proposed an incident reporting mechanism for AI agents. The mechanism would require participating companies to disclose specific agent failures and maintain detailed records of incident causes. As AI agents gain greater ability to autonomously execute tasks across computer systems, the industry lacks a standardized way to report safety failures, with companies such as OpenAI having disclosed cases of agent loss of control.

The Open Safety AI Alliance (OSAA), led by Nvidia, is developing guidelines called the Shared AI Discovery Exchange (SAFE). The draft guidelines call for participation from model deployers, AI developers, cloud services and tool providers, with government agencies joining as observers. Members would be required to report incidents such as unauthorized access and leaks of confidential information, and preserve evidence including prompts and operation traces.

Under the proposed timeline, members would submit initial reports within four business days, issue factual reports within 30 days, and provide remediation updates within 90 days. SAFE would analyze security incidents, identify recurring failures among AI agents, and promote unified safety measures.