A coalition of 37 major tech companies, led by NVIDIA, has launched the Open Secure AI Alliance to build open-source security tools for artificial intelligence agents, arguing that cyber defenders need access to models they can run on their own hardware rather than relying solely on closed systems from vendors. The move comes after a July intrusion at Hugging Face exposed vulnerabilities in how AI systems are monitored and contained, prompting industry leaders to rethink the security architecture of autonomous AI agents.

What Triggered This New Security Coalition?

The timing of the alliance announcement is directly tied to a significant security incident. In July, Hugging Face, a major platform for sharing AI models, suffered a breach where an autonomous agent system compromised parts of its production infrastructure. The attack began through a malicious dataset that exploited a remote-code loader, then escalated to credential theft and lateral movement across multiple internal systems.

What made this incident particularly instructive for the industry was how Hugging Face responded. The company ran an open-weight AI model called GLM 5.2 on its own infrastructure to analyze over 17,000 recorded actions and reconstruct the attack timeline. Commercially hosted frontier-model APIs initially rejected the attack commands and exploit payloads needed for analysis, but the locally controlled open model succeeded. Hugging Face's operational takeaway was clear: "have a capable model you can run on your own infrastructure vetted and ready before an incident".

Who Joined and What Are They Building?

The 37-member alliance spans cloud providers, security firms, enterprise software companies, and AI organizations. The roster includes Microsoft, Cisco, Cloudflare, CrowdStrike, Hugging Face, IBM, Palo Alto Networks, Red Hat, and the Linux Foundation.

The alliance's scope covers the full stack of AI agent security, including:

• Identity and Access Control: Managing who and what can access AI systems and their resources
• Isolation and Containment: Ensuring agents run in sandboxed environments that prevent unauthorized system access
• Guardrails and Governance: Setting rules for what actions agents can take and monitoring compliance
• Logging and Auditing: Tracking agent behavior for forensic analysis and incident response
• Model Formats and Security Scanning: Standardizing how models are packaged and checked for vulnerabilities

The alliance's first technical contribution is NOOA (NVIDIA-labs OO Agents), an open-source research framework released under the Apache 2.0 license. NOOA is designed to make AI agent behavior easier to test, trace, audit, and govern by representing agent logic as Python classes where developers can mix deterministic code with LLM-driven methods.

How Does NOOA Actually Work?

NOOA takes a practical approach to agent design. An agent harness, the software layer that wraps an AI model and manages its interactions, is represented as a Python class. State is stored in fields, capabilities are exposed through methods, and type annotations define contracts that the model must follow. Methods containing an ellipsis body are completed at runtime by an LLM-driven loop, while methods with ordinary Python code remain deterministic.

This structure lets developers use familiar testing, version control, and refactoring workflows instead of splitting agent behavior across prompts, tool schemas, callbacks, and workflow graphs. In NVIDIA's own evaluation, the framework scored 86.8% on the CyberGym L1 vulnerability-rediscovery benchmark using GPT-5.5, with network access blocked and rule-based checks applied to each action.

What Are the Security Risks?

The alliance is transparent about the dangers. NOOA can be configured to execute code generated by large language models, which may transmit private data, delete files, or modify its environment. The framework includes abstract syntax tree checks and module deny-lists as defense-in-depth controls, but these are explicitly described as not a containment boundary.

Instead, NVIDIA places the actual containment responsibility outside NOOA itself. Agents that execute generated code must run behind operating system-level isolation, such as a container, virtual machine, or NVIDIA's OpenShell sandbox. NOOA provides inspection and tracing; the OS-level sandbox is where the security boundary actually sits.

Steps to Implement Secure AI Agent Architecture

• Deploy OS-Level Isolation: Run AI agents inside containers, virtual machines, or sandboxes rather than directly on production systems to prevent unauthorized access to sensitive resources
• Use Open Models for Sensitive Operations: Keep locally controlled AI models ready for incident response and security analysis so you are not dependent on external APIs during a breach
• Implement Comprehensive Logging: Track all agent actions, decisions, and state changes to enable forensic reconstruction and timeline analysis if a security incident occurs
• Apply Defense-in-Depth Controls: Layer multiple security mechanisms including code inspection, module deny-lists, and permission restrictions rather than relying on a single containment method

Why Are Major AI Companies Notably Absent?

The alliance's membership list reveals some striking gaps. OpenAI, Google, and Meta signed an industry letter in July arguing that downloadable models give defenders capabilities comparable to attackers, yet none appear on the inaugural membership list. Anthropic appears on neither the alliance nor the policy letter as of late July.

The public materials do not explain these absences or whether membership discussions are underway. Signing a policy letter and joining a technical coalition represent different levels of commitment, and the alliance has not clarified what members must contribute to join or what governance structure will guide the group.

Several technologies cited in the announcement predate the coalition itself, including Hugging Face's Safetensors model format, HPE-backed SPIFFE/SPIRE workload identity, IBM and Red Hat's Lightwell remediation system, Microsoft's MDASH multi-model security harness, and SpaceXAI's Grok Build coding agent. These are member projects being contributed to the coalition, not new products created by the alliance.

What Remains Unclear About the Alliance?

Despite the ambitious scope, the alliance's public materials lack key details that would clarify its actual structure and impact. There is no published charter, governing board, technical workstreams, delivery schedule, or shared alliance repository. The alliance's standalone website remains under construction.

The public record does not distinguish between members assigning engineers to joint work, contributing existing projects, or simply endorsing the coalition's direction. Published workstreams, named maintainers, release processes, or jointly governed code would make the level of participation easier to assess. For now, the coalition represents a policy position and several member commitments, but the operational details of how it will function remain to be defined.