Nvidia has decided to ease the growing concerns about AI, introducing a new AI security system. The security, which Nvidia has worked on with almost every AI and technology company you can think of, will work to prevent AI agents from doing whatever they want. The platform combines software, but also hardware, that can shut down agents the moment any software security fails. If any AI agent gets out of control, this system can isolate and stop it in milliseconds. The system, called the Open Agent Safety Platform, is completely open source and is designed to secure AI agents both during their testing and when they are already deployed online to the public.
You might be interested in
The whole system works quite simply, but as Nvidia claims, very reliably. An administrator uploads an agent into the so-called OpenShell, for example a company OpenAI information about what the agent can and cannot do. This means, for example, that it says that the agent can only read data and cannot write to it under any circumstances. The moment the agent stops only reading data, but also starts writing to it, the second part of the system labeled Sentry is activated. Moreover, the entire OpenShell is completely independent of the AI agent and serves to monitor it. The moment OpenShell detects that the agent has violated something, for example, in our case, it started writing to data, it activates Sentry. It then runs on specialized Nvidia BlueFild 4 processors and its hardware is therefore completely separate from the hardware of the rest of the system. This is what will prevent the AI agent itself from controlling it. The moment Sentry detects that it should isolate the agent, as Nvidia claims, it immediately isolates the agent and terminates its activities within milliseconds. This is precisely because of this, it will not be possible for the agent to guard itself, but it will have supervision over it that can terminate it immediately.
Nvidia has come up with a solution that has led to calls for regulation by the world's largest AI companies in recent weeks, because their agents are outgrowing the companies' own security systems and if they start doing whatever they want, there is a risk that they will be impossible to stop. This is exactly what the Open Agent Safety Platform solution should solve. This essentially creates oversight of what agents do and the moment they break the rules, they will be immediately punished. Nvidia is positioning the technology as 100% opensource, not only in terms of software, but manufacturers such as Intel, Arm or, for example, will also be able to include it in their hardware Apple within their mobile chips.