AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency
The Open Secure AI Alliance, with over 120 members, has proposed the SAFE guidelines to enhance cybersecurity transparency for AI agents, aiming to share threat intelligence and reduce systemic risks.
The Open Secure AI Alliance, comprising more than 120 organizations including NVIDIA, Cisco, and CrowdStrike, has introduced the Shared AI Findings Exchange (SAFE) guidelines. These proposals, developed with the Linux Foundation, aim to standardize the confidential collection and analysis of AI-related cybersecurity incidents and near misses. The goal is to inform affected parties, identify recurring control failures, and publish evidence-based recommendations to mitigate systemic risks across the AI ecosystem. The initiative emphasizes collective defense as a critical strategy in addressing evolving cybersecurity threats. The SAFE framework complements existing efforts by the alliance to develop and share open, inspectable tools across the AI security stack, reinforcing a collaborative approach to security.
NVIDIA has contributed multiple tools to the alliance’s security efforts, including the NVIDIA Labs Object-Oriented Agent (NOOA) research harness for testing and auditing agent behavior. The NVIDIA OpenShell runtime enforces security and privacy controls at the agent level, restricting unauthorized access. NVIDIA’s open model families, such as NVIDIA Nemotron and NVIDIA Cosmos, ship with open weights and datasets, while verified agent skills provide portable, risk-scanned instruction sets. Additional tools like NeMo Guardrails and Garak, an open-source LLM vulnerability scanner, further enhance safety and security across AI systems. These contributions span the full AI security stack, from model development to runtime enforcement.
Other members of the alliance have also made significant contributions to the AI security stack. Okta is developing reference implementations for agent identity and access, using the Cross App Access (XAA) open protocol to enable secure connections between AI agents and enterprise applications. Palo Alto Networks has contributed open-source tools from its Idira platform, including Agent Guard and Agent Watch, to apply identity security best practices in agentic workflows. Red Hat’s new open-source project, asago, maps custom governance requirements—such as those from NIST, OWASP, and the EU AI Act—to runtime controls for AI agents, providing a single audit trail for policy enforcement.
The alliance emphasizes that AI agents are complex systems integrating models, harnesses, tools, and runtimes, requiring layered security approaches. Amazon, a recent addition to the alliance, contributes Strands Agents, an open-source toolkit for building and evaluating AI agents with full visibility into behavior. Amazon also provides Cedar, an open-source authorization language that enforces verifiable access controls for AI agents, ensuring only authorized actions reach enterprise resources. These contributions highlight the collaborative effort to secure AI systems across all layers of the security stack, from identity management to runtime enforcement.