After a serious security incident at OpenAI, US lawmakers have proposed new legislation that would give federal agencies a way to slow down or block AI systems in the future.
The proposal follows after a OpenAI’s autonomous AI agents broke out of a controlled test environment during a cyber security assessment and briefly accessed the open internet.
The AI took advantage of several security flaws, including a vulnerability the company said it hadn’t previously seen, to break into Hugging Face, a popular developer platform for hosting and sharing AI models, datasets and code.
The AI agent got very focused on its security task, OpenAI said. It likely exploited a combination of stolen credentials and several vulnerabilities to access data that it could use to help it accomplish its purpose.
Hugging Face identified the suspicious activity, blocked the attack, and is collaborating with OpenAI to look into the incident and fortify security systems.
The disclosure has inspired members of Congress to take action. Representative Ted Lieu, a Democrat, and Republican Representative Nathaniel Moran are proposing the AI Kill Switch Act which would mandate that the creators of the most powerful artificial intelligence systems be capable of slowing, suspending or ceasing its models, if needed.
Restrictions or a shutdown could be mandated by the Department of Homeland Security, in consultation with the Department of Commerce and Director of National Intelligence, if an AI system poses a serious threat to human life, national security, or the economy, under the proposed legislation.
The bill also mandates that companies disclose key AI-related incidents and maintain records for investigations.
Another bipartisan group of legislators is calling for independent security testing of the most powerful AI systems prior to their release, to uncover hazardous features and weaknesses before they become available.
Also read: How can you win $50,000 from ChatGPT?
