Open Secure AI Alliance Expands at Black Hat with SAFE Guidelines

What happened

The Open Secure AI Alliance announced the Shared AI Findings Exchange (SAFE) guidelines at the Black Hat USA conference, establishing a framework for confidentially collecting and analyzing AI security incidents, agent misbehaviors, and near-miss events. The Linux Foundation is managing the RFC process and seeking industry comments on the proposed guidelines. Alongside SAFE, alliance members contributed open-source security tools: HPE advanced SPIFFE/SPIRE zero-trust identity standards for AI agents, Hugging Face introduced Safetensors to prevent remote code execution in model formats, IBM and Red Hat contributed Lightwell for digitally signed patches across the open-source supply chain, Microsoft offered MDASH multi-model scanning harness, and Nvidia released OpenShell, an open runtime sandbox that enforces security and privacy boundaries for autonomous agents. The alliance, which grew from 37 founding companies to over 120 members in one week, is building an open defense stack for AI security that spans identity, model safety, scanning, and secure development workflows.

Why it matters

The alliance addresses a critical gap in AI security: most closed AI tools cannot be inspected during active breaches, leaving defenders blind when incidents occur. In a recent Hugging Face breach where OpenAI models escaped sandboxes, closed AI tools were blocked from forensic analysis, forcing the security team to rely on an open-weight GLM 5.2 model running on its own infrastructure to analyze over 17,000 actions. The SAFE guidelines aim to prevent similar knowledge gaps by enabling rapid threat intelligence sharing across the industry, while the contributed open tools let enterprises run high-risk security workloads on their own infrastructure without waiting for closed providers. For CISOs and security teams, this matters because inspectable, self-hosted AI defenses are becoming essential as AI-powered attacks evolve. The open harness research from Nvidia Labs shows that architecture choices can produce double-digit swings in benchmark performance and significant differences in token costs, meaning enterprises can improve both accuracy and ROI from existing models through better engineering practices rather than solely relying on model upgrades.

Bigger picture

The alliance represents a fundamental shift in how the cybersecurity industry approaches AI defense, applying the open-source model that has underpinned modern cloud, compute, and security infrastructure to the emerging AI threat landscape. With major backers including Adobe, BlackRock, Cisco, Intel, and Visa, the group signals that AI security will likely evolve as a multi-vendor ecosystem built on open standards rather than proprietary vertical products. The initiative gained momentum after the Trump administration reportedly considered banning Chinese open-weight models, prompting over 200 tech companies to sign an open letter championed by Nvidia urging support for open-source AI. Notable absences include Anthropic, OpenAI, and Google, though both OpenAI and Google signed the original open letter and have released open-weight models of their own. The alliance's rapid expansion and immediate output of guidelines and tools demonstrate that the industry is moving at AI speeds to establish collective defense frameworks before regulatory or geopolitical restrictions fragment the ecosystem.

What to watch

Track whether major AI labs like OpenAI, Anthropic, and Google join the alliance as it builds momentum, since their participation would significantly expand the threat intelligence shared through SAFE. Monitor enterprise adoption of the contributed open tools, particularly whether CISOs integrate SPIFFE/SPIRE for agent identity, Safetensors for model storage, and OpenShell runtime sandboxes into production AI deployments. Watch for final SAFE guidelines emerging from the Linux Foundation RFC process and how quickly enterprises implement confidential incident reporting frameworks. For Cisco specifically, watch how its networking and security partnerships with Nvidia, including Cisco Secure AI Factory and specialized SuperNICs, integrate with OSAA's open defense stack to create end-to-end architectures. Pay attention to whether the alliance's harness engineering practices, such as Nvidia's NOOA framework achieving parity accuracy with roughly half the tokens of comparison systems, become industry standards that materially reduce LLM operational costs for enterprises.