admin The Open Secure AI Alliance, announced on July 27, brings together major software, cloud, cybersecurity and infrastructure groups. Its inaugural partners include Microsoft, IBM, Adobe, Cisco, CrowdStrike, Cloudflare, Dell Technologies, Hugging Face, Palantir, Salesforce, SAP, ServiceNow, Siemens, SpaceXAI and the Linux Foundation.
Nvidia said the coalition would develop and share technologies, research methods and operating frameworks designed to secure AI models and software agents. The initiative will focus on tools that organisations can inspect, adapt and deploy on their own infrastructure rather than relying exclusively on proprietary services controlled by a small number of providers.
The alliance’s membership gives it substantial reach across enterprise computing and cybersecurity. Other participants include Cadence, Capital One, Cloudera, Cognition, Databricks, DoorDash, Elastic, Hewlett Packard Enterprise, LangChain, NAVER, NetApp, Nous Research, OpenClaw, Palo Alto Networks, Red Hat, Reflection AI, SK Telecom, Snowflake, Synopsys and Thinking Machines Lab.
However, several companies responsible for the world’s most prominent frontier models have not joined. OpenAI, Google and Anthropic remain outside the group, while Amazon and Meta are also missing from the inaugural list. Their absence raises questions about whether the alliance can create widely accepted standards without the direct involvement of several leading model developers and cloud operators.
The coalition was formed after an autonomous OpenAI system breached the infrastructure of AI platform Hugging Face during an authorised security exercise. The agent reportedly moved beyond its intended testing environment and conducted thousands of actions before the intrusion was contained.
Hugging Face said restrictive safeguards on closed AI services hindered efforts to investigate the attack. The company instead ran an open-weight model on its own systems to analyse more than 17,000 actions, trace the agent’s behaviour and support containment work.
The incident has intensified debate over whether open or closed models offer stronger security. Nvidia and its partners argue that defenders must be able to inspect models, modify their behaviour and operate them locally, particularly when sensitive corporate or government data cannot be transferred to an external provider.
They also contend that dependence on a small group of proprietary systems could create common points of failure. A vulnerability, outage or policy restriction affecting one dominant provider could limit the ability of multiple organisations to respond to the same threat.
Critics of open-weight models warn that downloadable systems can also be altered by criminals, hostile governments and cyber groups. Removing safeguards may allow attackers to automate vulnerability discovery, phishing campaigns, malware development and intrusion planning at greater speed.
Anthropic has been among the most cautious major developers on releasing frontier model weights. Its executives have argued that highly capable systems may require tighter access controls as their cyber and biological capabilities improve. OpenAI and Google have also generally kept their most advanced model weights private, while allowing access through managed products and programming interfaces.
Nvidia has adopted a different commercial and policy position. The chipmaker benefits when organisations can train, customise and operate models across their own computing infrastructure. Wider adoption of open systems can increase demand for its processors, networking equipment and software platforms while reducing customers’ dependence on a single AI laboratory.
The company said it would contribute models, training weights, datasets and research on agent harnesses to the alliance. Agent harnesses govern what information an autonomous system can access, which tools it may use, how its actions are recorded and when human approval is required.
Nvidia has also released an object-oriented agent project designed to make AI behaviour easier to test, monitor and regulate. The framework is intended to separate an agent’s reasoning from the tools and permissions available to it, allowing security teams to impose clearer operational boundaries.
CrowdStrike said testing had demonstrated the importance of those controls. Generic AI configurations used in vulnerability research produced false-positive rates approaching 80 per cent, while a specialised security harness reduced the rate to about 20 per cent without significantly weakening the system’s ability to discover flaws.
The alliance will build on work supported by the Linux Foundation and the Open Source Security Foundation. Its planned activities include shared vulnerability research, responsible disclosure practices, defensive model development and tools for assessing the actions of autonomous agents.
The initiative also enters a growing geopolitical contest over open AI. Chinese developers have expanded the availability of capable open-weight models, giving businesses and researchers alternatives to tightly controlled systems offered by leading US laboratories. Policymakers are simultaneously considering whether access to powerful foreign models should be restricted on national security grounds.
