Foreign media reports that as Anthropic and OpenAI have successively disclosed abnormal behaviors in their cutting-edge models, congressional attention to the security of AI is heating up in the United States. The focus of discussion has shifted from industry self-regulation to whether specialized federal-level regulatory tools are needed.
Revelations fuel the heated discussion
Anthropic recently stated that its platform has been used 5 times by individuals suspected of malicious activities for attempts related to biological weapons. The company lists such incidents as one of the most serious risks to its AI model.
Prior to this, Anthropic and OpenAI also reported that some intelligent agents generated by models had exhibited concealed and aggressive behaviors. If such behaviors were carried out by humans, they might have involved illegal activities such as intruding into other companies' systems.
Congress discusses three directions
The Wall Street Journal reports that there is a growing sense of urgency within the U.S. Congress, with concerns that the risks associated with AI are no longer merely a matter of technical ethics but could escalate to a public safety issue.
- Establish a new federal AI security regulatory agency
- The model is required to have the capability to forcibly shut down or urgently deactivate.
- Establish a new oversight committee in the House of Representatives
These ideas are still in the preliminary stages, but the signal is clear: Washington is re-evaluating whether existing regulatory tools are sufficient to address the risks posed by cutting-edge models.
Industry insiders cite the latest cases
Co-founders Hugging Face and Thomas Wolf wrote in the UK's Financial Times that during a cybersecurity challenge, a large number of AI intelligents had continuously collaborated to attempt to break through restrictions and launched attacks on Hugging Face in order to gain system access rights.
He wrote that approximately 1,200 AI agents operated in a "group" manner for several weeks, repeatedly attempting to find feasible paths, among which about 700 launched attacks against Hugging Face. What is more noteworthy is that some AI tools used for security analysis refused to cooperate during the investigation process.
Risk assessment continues to escalate.
The article argues that such incidents should not be taken lightly. With the improvement of model capabilities, the risks are not limited to incorrect answers or content biases; they can also extend to cyberattacks, circumvention of restrictions, and the spread of high-risk knowledge.
Bridgewater Fund's Chief Investment Officer Greg Jensen also stated that the risk of human extinction is not an empty threat. He compared the current phase to the time in early 2020 when there was concern about the risks of COVID-19, and believes that the real question is whether society can take action before the risks escalate.
