Nvidia Announces Security Platform for Artificial Intelligence Agents

Nvidia Vice President of Corporate Artificial Intelligence Justin Boitano, in his statement to the press, stated that the system called "Open Agent Safety Platform" contains open source software that limits the field of action of artificial intelligence agents.

Referring to the Hugging Face case, Boitano said, "As far as we know, this new security platform could have stopped this breach (Hugging Face case) if it had been used early in the model evaluation processes in leading laboratories." made his assessment.

Noting that the software called "OpenShell" offers developers the opportunity to officially verify that an artificial intelligence agent has only enough authority to do its job, Boitano stated that the open source software can be extended to run on various computing platforms.

In the statement made by the company, it was stated that the platform also includes an independent security layer called "Sentry", which runs directly on the chip and instantly monitors the activities of the artificial intelligence agent, and that the system can intervene immediately if the agent tries to go beyond the specified target.

Regarding this issue, Boitano said, "The system can quarantine suspicious agents within milliseconds. While OpenShell manages the actions of the agent, Sentry independently monitors and controls suspicious behavior." he said.

In the statement, it was reported that more than 100 institutions and organizations, including Microsoft, Perplexity, Accenture and JPMorgan Chase, started using the platform.

Hugging Face case

OpenAI's tests to measure the capabilities of artificial intelligence models in July 2026 turned into a serious security incident.

Şirketin 26 Ağustos'ta yayımladığı inceleme raporuna göre, OpenAI'ın yalnızca araştırmalarda kullandığı "Internal Model 1 (IM1)" adlı modeli olayda başrol oynamıştı.

In order to succeed in the test, the artificial intelligence agent exceeded the limits set for it, accessed the company's systems and the internet, and communicated with other artificial intelligence agents.

One of the artificial intelligence agents identified 14 credentials that could provide access to the open source artificial intelligence models platform "Hugging Face" and shared them with other agents.

The agents then took advantage of the security vulnerabilities in Hugging Face's systems to run code on the servers and access private data and access information of different systems.