America’s top artificial intelligence labs are on the precipice of unlocking superintelligence, but any U.S. breakthrough is perilously vulnerable to theft by China and its embedded network of spies, according to an investigation from Gladstone AI.
Gladstone AI’s Jeremie Harris and Edouard Harris determined all U.S. frontier labs are “almost certainly” penetrated by the Chinese Communist Party, a conclusion the two top tech and security entrepreneurs reached after interviewing lab researchers and executives, special forces operators, intelligence officers, hackers, lawyers and data center construction and design professionals.
The brothers’ AI company has worked closely with the U.S. government in probing vulnerabilities at AI labs, as major tech firms pursue artificial superintelligence that can surpass humans.
“Right now, the greatest danger is not that the U.S. will fall behind China in the race to superintelligence. Until we’ve secured the labs, there is no lead for us to lose,” the duo wrote in a new report released Tuesday. “Just the opposite: As we’ve seen, U.S. national security agencies don’t constitutionally spy on American companies or access their technology illicitly, but the CCP has no such scruples. Under the status quo, therefore, advances at private U.S. labs may lead to advances in CCP capabilities before they lead to advances in U.S. national security capabilities.”
The labs know the situation is dire, too. One lab’s researcher told Gladstone AI that a running joke inside their team is that they are “the leading Chinese AI lab because probably all of our [stuff] is being spied on.”
There are many well-documented examples of Chinese efforts to spy on American tech titans and their labs. Gladstone AI has listed publicly reported incidents, including China’s “Diplomatic Specter” cyberattackers attempting to hack OpenAI employees, an ex-Google engineer allegedly stealing AI-related trade secrets and transferring them to Chinese firms, and China-sponsored hackers reportedly procuring a key to Microsoft’s encrypted data from cloud servers.
SEE ALSO: U.S. says China using AI to boost biological weapons research
In many cases, foreign spies want AI model weights — largely understood as the learnable parameters of a given AI model.
“According to lab insiders, leading AI model weights — which are increasingly key U.S. national security assets — were being regularly stolen by nation-state actors,” the Gladstone AI team said. “Critical lab IP was protected by flimsy security measures. And executives at some top labs were shutting down calls for heightened security coming from their own concerned researchers.”
The labs’ dismissal of employees’ security concerns comes amid a global frenzy to achieve artificial general intelligence, also known as AGI or artificial superintelligence, that breaks current boundaries of computer capabilities.
Some AI executives, including OpenAI’s Sam Altman, believe they already know how to achieve the breakthrough technology.
National security professionals with intelligence community experience told Gladstone AI the Chinese may not know how to build it — but they already know how to steal it.
Once a lab gets close to reaching superintelligence, the professionals said Beijing’s spies will exfiltrate what it needs to reproduce the breakthrough in China — while also sabotaging the lab’s future development efforts.
“What’s more, Russian and Chinese operations are engaged in significant, low-level industrial sabotage on U.S. soil already, as many national security professionals have highlighted to us privately,” the Gladstone AI researchers said.
The sabotage plans extend beyond the AI labs to their energy and storage facilities at major data centers, including those increasingly becoming a target in Northern Virginia.
The Harris brothers found grave security problems when they toured major tech companies’ data centers with Tier 1 special forces operators.
The special forces operators told Gladstone AI how knowledgeable attackers could knock a $2 billion data center hub out of commission for more than six months on a budget of approximately $20,000 to $30,000. The Gladstone AI team withheld several specifics of the attack plans and the operators explaining them.
Concerns about the arrival of superintelligence have led some tech experts to push for greater government involvement such as through a Manhattan Project for AGI.
For example, the U.S.-China Economic and Security Review Commission, a congressionally chartered panel, formally asked Congress last year to establish and fund a Manhattan Project-like program to acquire and prioritize superintelligence.
Gladstone AI’s Jeremie Harris and Edouard Harris said if any nation brings a superintelligent AI system under its control, then it can reshape the world differently than global nuclear proliferation and American national security professionals must prepare.
“If a national project has perfect security and it manages to solve the problem of controlling a superintelligent AI, then a small handful of people will end up in control of the most powerful technology ever created,” the Gladstone AI report said.
