AI 安全與開源工具
Open Secure AI Alliance Launches, with NVIDIA Advancing Agent Testing, Tracing, and Auditing Tools Through NOOA
NVIDIA, the Linux Foundation, Hugging Face, Microsoft, and other organizations have formed the Open Secure AI Alliance to address vulnerabilities in agents and AI software with open models and tools. Its first concrete contributions include NVIDIA’s NOOA framework, but the alliance has yet to announce common specifications, governance processes, or a delivery timeline.
NVIDIA and dozens of cloud, cybersecurity, enterprise software, and open-source organizations have formed the Open Secure AI Alliance, with the goal of jointly developing inspectable, self-hostable AI defense tools. Its initial members include the Linux Foundation, Hugging Face, Microsoft, GitHub, IBM, Red Hat, Cloudflare, CrowdStrike, LangChain, vLLM, and numerous model and infrastructure providers. The alliance will build on the Linux Foundation’s Akrites project and OpenSSF’s software supply chain security work, extending vulnerability discovery, disclosure, and remediation processes to models, agent runtimes, and related software.
The most concrete technical contribution so far is NVIDIA’s release of NOOA for general use. The Linux Foundation describes it as an open-source framework for testing, tracing, auditing, and governing AI agents. The alliance also plans to explore open weights, data, and agent harnesses, enabling defenders to replay behavior on their own infrastructure, inspect tool calls, and adjust controls. This direction addresses practical constraints in agent security incidents: when responders have only API access, they often cannot inspect the model’s internal version, pin the inference environment, or reproduce the complete attack trajectory in an air-gapped network.
For engineering teams, the alliance’s real value will depend not on its membership roster but on whether it can deliver cross-vendor incident formats, identity and authorization models, reproducible test harnesses, and interfaces compatible with OpenTelemetry, SBOMs, and vulnerability disclosure processes. For now, the announcement remains largely a mission statement, with no stable APIs, threat model, version roadmap, or neutral technical governance mechanism specified. NOOA’s licensing, deployable components, and integration maturity will also need to be assessed based on the actual repositories. Key issues to watch include whether deliverables are placed under foundation governance, whether the work can cover both closed-source APIs and open-weight models, and whether alliance members publish reproducible benchmarks for agent attacks and defenses.