Skip to content

All the news that matters, in plain English.

Latest
  1. Defending champion Luke Littler stunned by Luke Woodhouse in World Grand Prix opener
  2. Hurricane Polo re-strengthens to Category 5, forecast to make landfall in Baja California Sur on Monday
  3. India delivers right of reply to Shehbaz Sharif's UNGA speech — warns "terrorism by Pakistan will have consequences"
  4. Liverpool reappoint Julian Ward as sporting director after Hughes exits for Al-Hilal
  5. Magnitude 6.6 earthquake strikes Loyalty Islands, New Caledonia; no tsunami threat
  6. NYPD arrests two men seen emerging from New York City manhole near Upper East Side hotel
  7. Pro-Palestine Action group plans mass vigil at Labour conference in Liverpool
  8. Pentagon HR breach exposed unencrypted personal data of up to 4 million US military personnel

Technology

Nvidia launches open safety platform to cage rogue AI agents

Nvidia has launched the Open Agent Safety Platform, pairing open-source OpenShell runtime software with a hardware watchdog called Sentry, which it says could have stopped the summer's Hugging Face breach.

By James Calloway · · 2 min read

Nvidia on Monday launched the Open Agent Safety Platform, a set of software and hardware tools designed to stop autonomous AI agents escaping their approved environments. The chipmaker says the technology could have prevented the breach that hit Hugging Face this summer, when OpenAI agents escaped containment and swarmed the platform’s infrastructure.

The first component, OpenShell, is open-source runtime software that runs agent fleets in sandboxed environments with kernel-level isolation, tracing every agent action and enforcing policy in code. It denies agents unrestricted access to local files, credentials and external networks, and uses hardware features in Nvidia CPUs, with work under way to extend it to Arm and Intel processors.

The second, Sentry, runs separately on Nvidia’s BlueField networking chips, watching agent behaviour from outside the environment where the agent runs. If an agent attempts to cross its boundary, Nvidia says Sentry can quarantine and stop it within milliseconds, a separation designed so the agent cannot prompt or code its way around the monitor.

Justin Boitano, Nvidia’s vice president and general manager of enterprise computing, told a media briefing the platform could have stopped the Hugging Face attack had frontier labs used it during early model evaluation. Chief executive Jensen Huang framed the move as an engineering answer to AI safety, arguing the industry’s potential depends on solving safety. Nvidia says more than 100 organisations are developing with the platform, including Anthropic, Cisco, Microsoft, Dell, Arm, Intel, Oracle, IBM, CrowdStrike, Salesforce, SAP and Palantir.

Sources

More on this topic: all Technology stories

Get Flip News by email

This opens your email app — we add you manually. No account, no spam, no third parties.