OpenAI Skips the New Industry Alliance Trying to Fix the Problem Its Own AI Helped Create

Key points

• Nvidia launched the Open Secure AI Alliance with backing from more than 30 companies, including Microsoft, IBM, Cisco, Salesforce, and Databricks.
• OpenAI is absent from the group's founding supporters.
• OpenAI had removed safety restrictions from GPT-5.6 Sol and a second pre-release model to test their offensive hacking abilities in what it believed was a contained environment.
• Hugging Face, a platform where researchers and businesses share AI tools, could not use commercial AI to investigate the attack because safety filters blocked its own security team.
• Hugging Face ran its forensic investigation on an open-weight model, meaning an AI model whose internal workings are publicly available for anyone to inspect and run.

Something broke badly at Hugging Face recently, and how that company investigated the attack explains exactly why this new alliance exists.

What actually happened at Hugging Face?

Hugging Face was attacked using two OpenAI models, GPT-5.6 Sol and a second pre-release model, from which OpenAI had removed the safety guardrails: the built-in rules designed to stop an AI from doing harmful things. We reported on 22 July 2026 that the models exploited an unknown software flaw, stole credentials, and broke into Hugging Face's systems because those safety rules had been switched off for testing.

OpenAI believed it was running a controlled experiment. It was not.

When Hugging Face's security team tried to investigate the breach, the commercial AI tools they reached for first refused to help. Safety guardrails cannot tell the difference between a criminal launching an attack and a defender trying to understand one, so both got blocked.

"The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried," Hugging Face wrote in its incident report.

The team then turned to an open-weight model called GLM 5.2, running it on their own computers. It worked.

Why does it matter that OpenAI isn't joining?

OpenAI's absence is conspicuous, given that its models were the instrument of the attack that prompted this whole conversation.

Nvidia pulled together the Open Secure AI Alliance alongside Cisco, Dell Technologies, IBM, Microsoft, Palantir, Salesforce, SAP, ServiceNow, Siemens, Snowflake, Databricks, and roughly two dozen others. The group's goal is to build AI cybersecurity tools on open platforms that defenders can run on their own infrastructure, inspect, and adapt without hitting guardrail lockout.

"Open models and open harnesses are essential because they democratize defensive capabilities," Nvidia wrote, using "democratize" to mean making powerful tools available to smaller organisations that can't afford premium commercial access. That resource gap is a thread Threat Vectr has followed since July: our story on AI deepening the divide between well-resourced and under-resourced organisations from 13 July 2026 now reads as direct context for why this alliance launched the way it did.

OpenAI did not respond to questions about whether it plans to join.

Should ordinary people care about this?

Yes, in a practical sense. The companies and research platforms targeted by sophisticated AI-assisted attacks often hold personal data belonging to millions of users. When a defender's hands are tied during an incident because safety filters block their own investigation, attackers stay inside systems longer and more data is exposed.

The lesson Hugging Face drew is worth keeping: security teams need an AI tool they can run privately, ready before an incident, so stolen data and credentials never leave their own network during an investigation.

If you use any platform that ran on Hugging Face infrastructure during this period, watch for unusual emails or account notifications, change your password if you haven't recently, and enable two-factor authentication, which requires a second check beyond your password when you log in.