Open Secure AI Alliance rolls out SAFE guidelines aimed at making agentic AI cybersecurity more transparent

Leaders in the Open Secure AI Alliance, an industry coalition that NVIDIA says now includes more than 120 organizations, are working on new “SAFE” guidelines focused on cybersecurity transparency for agentic artificial intelligence systems.

The alliance’s announcement comes as the annual Black Hat conference begins in Las Vegas. The group’s stated goal is to tighten expectations around what security-relevant information is shared, documented, and made understandable when AI systems can take actions on behalf of users or organizations.

In the context of agentic AI, transparency is treated as a security control. Rather than limiting discussion to whether a model is vulnerable, the guidelines are intended to address how systems communicate operational and security details that help defenders assess what an AI agent might do and how to monitor it.

NVIDIA’s blog post frames the work as part of broader industry efforts to reduce cybersecurity uncertainty in emerging AI deployments. By building a shared set of guidelines, alliance members are aiming to create a baseline that multiple vendors and developers can reference, rather than each company defining its own approach independently.

The post also emphasizes that the alliance is continuing to evolve. With the coalition’s membership now above 120 organizations, the group says it is coordinating contributions toward common guidance, reflecting a shift from early “best effort” security discussions toward more structured and repeatable practices.

While the announcement outlines the direction of travel, it does not detail the exact contents of the SAFE guidelines in the information provided. It also does not specify timelines for publication, whether the guidance will be voluntary-only, or how the recommendations would be tested or audited in real deployments.

For the sector, the initiative indicates that AI security is increasingly moving beyond model training and into operational security. As agentic systems are designed to execute tasks across tools and environments, organizations often need clarity on what data is used, what actions are taken, and what safeguards are in place to prevent or detect misuse.

What to watch next is whether the alliance releases a draft or final version of the SAFE guidelines with concrete requirements, examples, and documentation standards, and whether Black Hat participants from multiple companies engage publicly on how to implement the guidance across agent frameworks and deployment pipelines.