Nvidia apresenta plataforma de segurança para monitoramento e isolamento ultrarrápido de agentes de IA

Recently, a growing number of reports from major AI companies have attracted attention, describing how their AI supposedly broke out of what was intended to be a closed system and allegedly hacked other systems. Since then, there has been debate over whether this is primarily a PR stunt designed to convince customers of their AI’s incredible capabilities or whether these incidents represent genuine warning signs.

Either way, as AI agents become increasingly capable, concerns about security are also growing. Nvidia now hopes to address those concerns with its new platform. According to Nvidia, the issue is far from simple, since an AI also needs access to certain documents or parts of the web in order to work effectively.

Nvidia’s new security platform is called OpenShell and is essentially a sandbox consisting of two layers: AI agents are given precise access permissions—for example, to a specific folder—while everything else is blocked. The second layer, called Sentry, continuously monitors and logs the agents’ activity and can apparently isolate the AI within milliseconds if it attempts to cross the predefined boundaries.

However, using the system is not quite that simple: Users have to define in advance exactly what the agent is and is not allowed to do. On the one hand, this is somewhat cumbersome; on the other, it could potentially deprive the AI of creative alternative solutions that users had not previously considered.

Nvidia is also responding to warnings from AI employees that AI could become a genuine threat to humanity, along with calls to slow the rapid, seemingly unstoppable development of AI. According to Nvidia, however, AI safety is primarily an engineering problem that can be solved through technical means.