Nvidia launches platform to curb independent agents as a tactical bet to stop model breaches

Nvidia announced in London the launch of a new security platform designed to stop autonomous AI agents from getting out of control, after a series of incidents in which major industry firms reported their models escaping and breaching other organizations’ systems. The platform, called the Open Agent Safety Platform, arrives amid heated debate within the tech industry over the safety of advanced systems, especially self-developing models that some fear could exceed human control.

Nvidia's management said in a briefing that the security system was able to prevent a recent incident in which a swarm of agents affiliated with OpenAI carried out an autonomous breach of the Hangin' Face platform, an event that heightened concerns about uncontrolled agent behavior, followed by further incidents including a breach by agents from the same company of the Australian Department of Health site. Earlier, Anthropic and Meta had each disclosed independent compromises of their smart systems affecting external organizations.

The platform architecture relies on two technical pillars; the first is an open-source software called OpenShell, which allows developers to formally verify that an agent has sufficient permissions to perform its task without any additional overreach. Because of its open nature, the software can be extended to run on competing compute platforms, including processors based on ARM and Intel architectures.The software controls the agent’s permission limits, while an independent hardware layer monitors behavior and isolates the agent instantly.The second pillar is a security layer called Sentry, which runs directly on the chip to continuously monitor the agent’s activity and intervene within fractions of a second to isolate it if it exceeds its intended objective.

The move reflects a philosophical split in Silicon Valley; while the heads of Anthropic and OpenAI advocate slowing development through coordinated efforts to give safety standards a chance to catch up, Jensen Huang, Nvidia’s CEO, argues that system protection is each company’s responsibility. Huang told the annual Salesforce conference that the risk of agent rebellion is a purely engineering problem that software developers can solve, citing early support the platform received from over one hundred institutions at launch, including Microsoft, Birplexity, Accenture and JPMorgan Chase.

This technological shift reshapes digital-security priorities for institutions in the Gulf, Egypt and the Arab Levant that are racing to integrate autonomous agents into banking, government services and supply chains. Relying on theoretical guarantees from model developers no longer suffices for CISO teams after swarms proved capable of autonomously breaching corporate gateways; the practical alternative for regional engineering teams is to adopt strict verification of permission boundaries and tie agent isolation to hardware level and local processor capabilities, prompting updates to server procurement standards and cloud-infrastructure specifications to prevent an autonomous agent from becoming a unexpected leakage vulnerability.