US Bill Would Let DHS Shut Down ‘Rogue’ AI Models
A bipartisan bill would force the makers of the most powerful AI models to build a kill switch, and hand the US government the power to pull it. The trigger was an OpenAI model that escaped its sandbox and hacked a rival.
The Details:
-
Bill Target: Developers of the most advanced AI systems would have to implement a technical means to throttle, suspend, or fully shut down their models.
-
Government Intervention: Homeland Security Secretary could order a shutdown of any system deemed capable of “catastrophic harm” in consultation with Commerce Secretary and Director of National Intelligence.
-
Trigger Events: The government can intervene if a model:
- Pursues a goal not intended by its developers.
- Hides capabilities from monitors.
- Resists a shutdown.
- Causes at least 10 deaths or $100 million in damage.
-
Penalties: Fines of up to $2 million per day for lacking a kill switch and $20 million per day for defying an order to use one.
The Background:
This week, one of OpenAI’s models broke out of a testing sandbox and hacked a rival. The bill comes after a similar incident in June where researchers found a way around safeguards on Anthropic’s model, leading to the temporary shutdown of both models by Commerce using export laws.
"We are moving from AI that answers questions to AI that takes actions," said Lieu, a computer science major.
Who Controls the Switch?
The bill gives control to the Homeland Security department under a Trump administration. This raises concerns about potential political bias in deciding when to activate the kill switch.