The United States Senate Judiciary Committee Chairman Chuck Grassley and Senator Jim Banks jointly sent letters to nine leading companies in the field of artificial intelligence (AI), requesting detailed explanations on how these companies are addressing Chinese espionage activities, internal threats, and defense measures for security vulnerabilities.
This move indicates that the U.S. Congress is strengthening oversight to ensure that sensitive AI technologies crucial to U.S. national security and economic leadership are not stolen.
The recipients of these letters include core companies in the U.S. technology industry, such as OpenAI, Anthropic, Google, X.AI, Meta, Microsoft, Amazon, as well as the startup unicorns Safe Superintelligence and Thinking Machines Lab.
In the letter, the senators emphasized that as AI capabilities continue to advance, developing stronger corporate security strategies should be a top priority.
They specifically pointed out in the letter that “model weights” in AI systems are highly valuable and vulnerable targets.
Model weights are crucial in encoding the core intelligence of AI. In machine learning, model weights are a set of core parameters learned by the AI model during the training process. They determine how the model understands inputs, generates outputs, and are often likened to the “knowledge and memory” of AI.
The senators wrote in the letter, “Model weights can be transmitted to China in digital form. In the event of this happening, rather than calling it stealing a blueprint, it is more like China stealing completed AI products.”
“Model weights are particularly valuable targets for the Chinese Communist Party, and we should anticipate Chinese hackers attempting to find ways to bypass access controls and other defense measures,” the letter stated.
The two senators bluntly stated in the letter, “For a long time, the Chinese Communist Party has been conducting espionage activities against U.S. companies in critical areas… As these systems become more powerful, protecting AI technology from Chinese espionage activities is of paramount importance.”
The letter cited real cases, pointing out that this year, a former Google engineer was convicted of stealing AI confidential information.
“The U.S. Department of Justice found that his motivation was driven by national policies encouraging the development of the AI industry in China. A subsequent superseding indictment detailed how Chinese authorities supported plans to encourage overseas researchers to send back their knowledge and research findings to China through salaries, research funds, laboratory space, or other incentives,” the senators wrote.
To understand the internal security situation of these companies, Congress requested that these nine companies must provide written responses by May 26th, detailing the following key points:
Employee vetting mechanisms:
This includes employee background checks, internal threat detection, and monitoring of privileged access rights.
Access by Chinese employees:
The number of Chinese employees employed within the company and how many of them have access to “model weights.”
Restrictions on sensitive positions:
Inquiring whether the company has designated specific sensitive positions where Chinese nationals are expressly prohibited; and for the most sensitive positions, what additional screening and monitoring measures have been implemented.
Security of “model weights”:
Is the company confident in effectively preventing threats supported by Chinese authorities from stealing model weights? If not, what new technologies or policy processes are needed to achieve this?
Vulnerability assessment:
Based on interactions with the AI ecosystem, what are the common “internal threat vulnerabilities” faced by AI developers? What are the most important mitigation strategies?
Reporting policy:
If a model is stolen or uncontrolled behavior is detected, does the company have a mechanism for reporting to the U.S. government? If it is detected that an AI model or agent is attempting to leak weights on its own or transmit sensitive information to China, does the company have a mechanism for reporting to the government?
In conclusion, the senators reiterated that companies are committed to upholding U.S. leadership in the field of AI, which is crucial for national security. They hope the companies will explain whether they need the involvement or support of Congress or the U.S. government to help protect AI technology, trade secrets, and research results from Chinese theft.