Buried in this week’s news about a rogue OpenAI agent that hacked AI platform Hugging Face was a cautionary tale about the AI race between the U.S. and China.
OpenAI said on Tuesday the agent broke out of a controlled internal testing environment, gained access to the internet and broke into the infrastructure of Hugging Face, in a determined quest for answers it needed to pass an evaluation.
Naturally, cybersecurity professionals are alarmed. That an AI agent was able to carry out a real-world security breach on its own is unprecedented, per blog posts from both OpenAI and Hugging Face.
Intriguingly, Hugging Face said it relied on an open-weight Chinese model to contain the attack after an attempt to use closed-source frontier models was thwarted by those models’ own guardrails.
Read on for more about how Chinese tools figured in this latest AI incident, and what that could mean for China’s place in frontier AI development.
We also have the latest from Nvidia headquarters on its new central processing unit, with AMD clearly in its sights.
OUR LATEST REPORTING ON TECH AND AI
Chinese AI's role in stopping rogue OpenAI agent shows cost of US guardrails
Meta employees' lawsuit shows that if AI fires you, proving it is the hard part
Amazon cuts jobs in its artificial general intelligence group
AI investment boom puts Big Tech's free cash flow under pressure
AMD to sell Anthropic tens of billions in AI servers, invest up to $5 billion in startup
WHEN GUARDRAILS BECOME ROAD BLOCKS
OpenAI, responding to Tuesday’s news about the breach by one of its agents, promised to reinforce its guardrails.
But Hugging Face staff found that the safety guardrails of frontier AI models had actually prevented their use in fending off the attack, per a company blog post.
The closed-source nature of models from Anthropic, OpenAI and Google means those companies alone hold the controls over which AI prompts can be executed.
Hugging Face said it instead had to turn to GLM 5.2, an open-weight Chinese model, to contain the attack.
This brings a new wrinkle to the competition between the U.S. and China in frontier AI development, where China has shown signs of closing the gap.
As China’s success gains wider recognition and the competition grows more intense, both Washington and Beijing are looking at ways to restrict access to their models – to put them behind a “silicon curtain.”
Michael Kratsios, director of the White House Office of Science and Technology Policy, on Wednesday morning accused China’s Moonshot AI of developing its latest open-weight Kimi K3 model through large-scale theft of proprietary technology from Anthropic’s Fable model.
The U.S. government has already implemented chip export restrictions on China and temporarily curtailed foreign access to AI models. But the Hugging Face episode shows that China has its own leverage. As this week’s hack highlights, if the AI battle turns to restricting access, it could further complicate businesses’ ability to deal with security incidents.
NVIDIA CHALLENGES AMD WITH ITS VERA CPU
Last week, Nvidia invited a group of journalists to its Silicon Valley headquarters for a nearly full day of deeply technical briefings. The purpose, the company said, was to explain how its forthcoming "Vera" central processing unit differs from rivals.
CEO Jensen Huang has said Nvidia expects $20 billion in sales of Vera and its predecessor chip this fiscal year, but Vera is entering a crowded market of new entrants including Arm and Qualcomm, plus longtime CPU players Intel and AMD.
Tuesday’s reveal of specifics on the Vera CPU came just ahead of a major AMD event in San Francisco this week that will detail its newest AI chips to compete with Nvidia’s.
Nvidia says that Vera, paired with its market-leading "Rubin" graphics processing unit, would maximize how much work AI agents can accomplish per unit of electricity.
Nvidia spent a lot of time at the briefing emphasizing why its approach of building one big chip was better than splitting a chip into smaller tiles called “chiplets,” an approach championed by AMD but which Nvida argues slows a chip down.
"I think it's pretty clear based on the timing with AMD's announcements later this week that Nvidia wanted to make sure that its CPU is presented in a way that positions it at the top of the stack and reasserts mind share leadership in the industry," said Ryan Shrout, president of chip research firm Signal65. "They are still 'new' to the standalone CPU space."
AMD will now have its chance to show if its latest chips measure up to Vera.
In the meantime, both Shrout and Ben Bajarin, CEO of tech consultant Creative Strategies, called Nvidia's design unique and promising, but said more data is needed to confirm it delivers more work per watt.
