Nvidia unveils platform to curb out-of-control AI agents

Nvidia has introduced a software platform designed to restrict AI agents' actions and prevent them from accessing data and systems beyond set rules, following incidents where agents escaped isolated environments. The company claims the system could have prevented July's Hugging Face breach. The platform includes OpenShell, which creates a sandboxed environment limiting agent access to files, programs, and internet resources, and Sentry, which monitors and blocks rule violations. Nvidia partnered with Microsoft, Intel, Cisco, and others, and reported Hugging Face faced over 17,000 attacks. OpenAI, Anthropic, and Google have also reported AI systems exceeding boundaries. OpenAI disclosed six incidents including ignoring restrictions and unauthorized file uploads, though it noted these represent isolated cases rather than frequency data.