AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency

Summary

Supported by the Linux Foundation, the Open Secure AI Alliance is developing SAFE guidelines for confidentially sharing AI cybersecurity incidents and intercepted attacks within an ecosystem of over 120 organizations.

Why it matters

These guidelines aim to reduce systemic risk by turning isolated agency-level AI attacks into collective defenses through threat intelligence sharing. By standardizing incident collection processes and publishing evidence-based recommendations, the initiative seeks to keep pace with the rapid evolution of AI attack surfaces.

Highlights

• NVIDIA, by providing the NOOA research framework and OpenShell runtime, contributes to enforcing strict security controls at the agent level.
• Open-source tools such as Garak (large language model vulnerability scanner) and NeMo Guardrails help detect data leaks, prompt injections, and enforce security policies.
• Alliance members like Okta (XAA protocol), Amazon (Strands Agents, Cedar language) and Red Hat (asago) offer identity, permission, and governance frameworks for secure agency workflows.

Critical take

While the SAFE framework promotes transparency, its reliance on confidential data sharing may encounter legal and privacy hurdles; the effectiveness of open tools also heavily depends on consistent adoption across different organizations.