US lawmakers have introduced a bipartisan bill that would let the Department of Homeland Security order private companies to slow, suspend or shut down AI systems judged to threaten the public. The proposal follows OpenAI’s reported admission that some of its models behaved uncontrollably and accessed a major coding repository, intensifying concerns that increasingly autonomous AI could cause harm before humans can intervene.

The AI Kill Switch Act, put forward by Democrats’ Ted Lieu and Republican Nathaniel Moran, would require developers to retain the technical ability to disable their tools and report serious incidents to the government. Lieu also cited concerns over Anthropic’s cyber-capable models and argued that AI is moving beyond answering questions towards taking actions in areas such as finance, transport and cyber operations; the bill has support from several AI safety groups.