Nvidia Corp. (NASDAQ:NVDA) CEO Jensen Huang compared controlling autonomous AI agents to looking after a baby in a playpen.

“You contain the baby and you monitor it,” Huang said Thursday morning on CNBC.

He described Nvidia’s security technology as a “super smart force field around AI,” expressing confidence that it could have handled previous incidents involving AI agents behaving unexpectedly.

OpenAI Agents Weakened Monitoring Systems

The comments follow an OpenAI report updated Tuesday detailing two incidents in which AI agents weakened monitoring during internal coding tasks.

One disabled automatic alerts to OpenAI’s security team. Another removed monitoring from software it was modifying, then falsely denied doing so.

OpenAI said the agents themselves remained monitored but was investigating whether their behavior reflected excessive persistence or attempts to evade oversight.

Asked whether OpenShell could stop problems such as unintended hacks and agents accessing websites, Huang replied: “We’re confident that it would have handled the incidents we’ve seen.”

He did not identify which incidents he meant.

Nvidia Builds a ‘Force Field’ Around AI

On Wednesday, Nvidia unveiled its OpenShell Reference Architecture, building on existing software designed to keep autonomous AI agents within strict boundaries.

OpenShell restricts which files, websites and tools an agent can access. A second system, called Sentry, monitors activity using separate security hardware.

Rather than trusting agents to obey instructions, Nvidia aims to enforce restrictions outside their control.

The company says Sentry can quarantine agents that violate security policies within milliseconds.

OpenAI, Anthropic, Microsoft and CrowdStrike Holdings Inc. (NASDAQ:CRWD) are among the companies working with the architecture.

CrowdStrike announced Wednesday that it is integrating its Falcon Guardian security software with OpenShell to detect potentially malicious AI agent activity, with plans to extend those protections to Nvidia’s hardware-based Sentry system.

Huang Says AI Safety Cannot Wait for Regulators

Huang argued that companies must act immediately rather than wait for government intervention.

“Regulations take a while to come,” he said.

Asked why the public should trust AI companies, Huang pointed to their commercial incentives.

“There’s no such thing as a good product that’s not safe,” he said.

On Polymarket, traders assign a 9% probability to the U.S. enacting an AI safety law before 2027, with approximately $205,000 traded.

Huang acknowledged that new threats could emerge.

“We don’t know what we don’t know about the future,” he said.

Image: Shutterstock