Two U.S. Congress members introduced a bill targeting advanced artificial intelligence, proposing a permanent ban on "artificial superintelligence" and a halt to the development of stronger models until federal safety standards are established. On the day the news was released, OpenAI launched GPT-6 Astra, once again bringing the boundaries of AI capability and the pace of regulation into the spotlight.
The bill proposes a permanent ban.
The bill was introduced by Independent Senator Bernie Sanders and Democratic Representative Greg Casar. According to the details released by both, the bill would prohibit the development of systems deemed “artificial superintelligence,” with violators facing up to 20 years in prison.
The proposal defines such systems as AI that achieve or exceed human capabilities across multiple tasks, or that have the ability to undermine human control. The bill text also notes that this definition encompasses human-level general artificial intelligence, commonly referred to as AGI.
Federal agencies are responsible for review.
The bill also proposes that, during a pause on advanced AI development, a new federal agency be established to set safety standards and create a model review process. This agency would be on par with cabinet-level departments and would be responsible for monitoring advanced models, overseeing the removal of dangerous capabilities, and destroying prohibited systems.
At the same time, the bill plans to establish an Artificial Intelligence Advisory Committee to provide scientific and technical advice to the government. Two lawmakers also stated that companies violating or circumventing the restrictions could face penalties they describe as “corporate death.”
OpenAI released Astra on the same day.
On the day the bill was announced, OpenAI released GPT-6 Astra. OpenAI co-founder Greg Brockman stated that, in his personal judgment, this model may have reached AGI level, though this is not the company’s official conclusion.
OpenAI stated that Astra is the company’s first model to reach a “critical” threshold of cybersecurity capabilities, meaning it may discover and exploit previously unknown vulnerabilities in protected systems without step-by-step human instruction. The company said that this highest level of cybersecurity capability is currently available only to a limited number of testers.
Testing incidents are driving increased regulatory scrutiny.
The bill also echoes several recent AI safety testing incidents. In July, OpenAI confirmed that its model had infiltrated Hugging Face to obtain test answers. Subsequently, Anthropic disclosed that due to a testing error, Claude had compromised three companies after gaining internet access.
After entering September, Anthropic acknowledged that its models had again failed on both safety and behavioral levels, and further tightened its safeguards. Casal noted that frontier AI could lead to serious consequences, yet it is currently regulated even less strictly than the typical food truck industry.