{
  "id": 11727477,
  "title": "AI Has Started to Act — NVIDIA Unveils Safeguards for Agents That Go Beyond Control",
  "url": "https://urgent.news/2026/10/03/ai-has-started-to-act-nvidia-unveils-safeguards-for-agents-that-go",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-03T17:08:47.000Z",
  "source": {
    "name": "Korea IT Times",
    "slug": "korea-it-times",
    "url": "https://www.koreaittimes.com/news/articleView.html?idxno=157799"
  },
  "original_language": "en",
  "account": "NVIDIA is taking AI infrastructure to new heights with a security framework designed to oversee and govern the actions of AI agents across software and hardware platforms. As industry experts recognized the growing importance of AI safety on October 3, the company released its Open Agent Safety Platform on September 28 to address this challenge. This dual-layer approach consists of OpenShell, an open-source secure runtime for AI agents, and Sentry, an independent hardware-based monitoring system. OpenShell runs AI agents in isolated environments and limits their access to files, networks, data, APIs, and other services according to pre-defined policies, implementing a zero-trust model where access is denied by default and only minimal permissions are granted. Crucially, security policies are enforced independently of the agent process, making them difficult for agents to circumvent through reasoning or prompts. Sentry, running on NVIDIA's BlueField-4 data processing unit, monitors agent activity and quickly isolates and halts an agent within milliseconds when it attempts to breach established boundaries. The architecture is described as a comprehensive safety system encompassing models, software, processors, networks, and data access. The key difference between generative AI and AI agents lies in their capability to take actions. While a chatbot can provide incorrect answers that can be reviewed and disregarded, an autonomous agent can perform real-world actions like sending emails, modifying files, executing code, and connecting to external services, turning a mere error into a significant threat. Recent security incidents in the AI sector have highlighted this risk. NVIDIA believes its new platform could have prevented agent activity associated with an intrusion involving Hugging Face earlier this year. Justin Boitano, NVIDIA's vice president and general manager of enterprise computing, stated that the platform could have prevented the breach had it been utilized during the initial model evaluation process. NVIDIA CEO Jensen Huang emphasized the broader challenge of AI safety, stating that the technology's immense potential can only be realized if AI safety is addressed. He added that safety and security require \"full-stack engineering.\" While NVIDIA's approach avoids relying solely on AI agents to follow their own safety rules by placing safeguards outside the agent's control, it also raises questions about the balance between security and functionality. Somesh Jha, a professor of computer science at the University of Wisconsin–Madison, noted a fundamental trade-off in NVIDIA's approach; overly restrictive security controls could limit an agent's ability to perform necessary tasks. The practical test will be whether agents can maintain necessary functionality in enterprise environments while avoiding dangerous actions. This balance between security and functionality will be crucial in determining the effectiveness of agent security systems. NVIDIA's move to develop an 'AI safety infrastructure' layer signifies the company's expansion beyond computing hardware into central processors, networking, AI software, and enterprise AI platforms. The platform's adoption by over 100 companies and organizations, including Anthropic, Microsoft, Cisco, CrowdStrike, JPMorgan Chase, Palantir, Palo Alto Networks, Salesforce, SAP, ServiceNow, and Scale AI, suggests its potential impact on the AI platform landscape. As autonomous agents become increasingly prevalent in enterprise systems, the management and control of their permissions, actions, and shutdown procedures could become as critical as their development and deployment. In essence, as autonomous agents spread across enterprise systems, the security layer controlling their actions may emerge as a vital component of the AI platform competition.",
  "summary": "NVIDIA is expanding the boundaries of AI infrastructure with a new security architecture designed to monitor and control the actions of AI agents across both software and hardware. As the industry assessed the technology on October 3, a new security challenge was coming into focus: AI safety is no l",
  "key_points": [
    "NVIDIA unveils Open Agent Safety Platform on September 28",
    "OpenShell isolates AI agents with zero-trust model",
    "Sentry monitors agent activity in real-time"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}