The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

The Open Secure AI Alliance, comprising more than 120 organizations, unveiled the Shared AI Findings Exchange (SAFE) guidelines at the Black Hat conference in Las Vegas. Developed with the Linux Foundation, SAFE aims to standardize confidential reporting of AI-related cybersecurity incidents and near misses, enabling collective defense against emerging threats. The framework includes mechanisms to inform affected parties, identify recurring vulnerabilities, and publish evidence-based recommendations to reduce systemic risks across the AI ecosystem.

NVIDIA contributed multiple tools to the initiative, including the NVIDIA Labs Object-Oriented Agent (NOOA) research harness for testing agent behavior, the NVIDIA OpenShell runtime for enforcing security controls at the agent level, and open model families like NVIDIA Nemotron and NVIDIA Alpamayo. The company also provided open-source agent skills with cryptographic signing and risk scanning, along with tools such as NeMo Guardrails and Garak for safety policy enforcement and vulnerability assessment.

Other alliance members contributed complementary technologies across the AI security stack. Okta introduced reference implementations for agent identity and access using the Cross App Access (XAA) protocol, while Palo Alto Networks contributed open-source tools like Agent Guard and Agent Watch for identity security. Red Hat launched asago, a project mapping governance requirements from frameworks like NIST and the EU AI Act to runtime agent controls with a single audit trail.

The SAFE guidelines and associated tools address the growing complexity of AI agents, which integrate multiple models, harnesses, and tools. Alliance members emphasized that securing these systems requires layered defenses and open collaboration. Amazon joined the alliance, contributing Strands Agents for building and evaluating agentic systems and Cedar, an open-source authorization language for enforcing fine-grained access controls on AI agents.