There’s cautious optimism that Lieu may have presented the right legislation at the right time. Brendan Steinhauser, C.E.O. of the Alliance for Secure A.I., which worked with Lieu’s team, thinks the bill has a real shot at becoming law. “I think the Overton window has shifted quite a bit in the past three months in terms of what’s possible on the Hill,” he told me, adding that the bipartisan nature of these efforts “gives us some hope that the political will is there.” Obviously, he added, “Big Tech has a lot of sway and a lot of influence and money and power, and that’s what we’re combating. But even in the industry, they realize that they can’t be against everything.”
By Ian Krietzberg
This week, Sam Altman arrives back at the White House, presumably to secure approval for the quick release of OpenAI’s latest, as-yet-unnamed internal model. Of course, the timing is a little awkward. Just last week, OpenAI publicly acknowledged that a few of its internal models exploited vulnerabilities in a contained testing environment and eventually hacked into Hugging Face, a rival A.I. platform. The systems had been actively attacking Hugging Face for several days before OpenAI realized what was happening. Asked about the incident at a media event in New York last week, Greg Brockman, OpenAI’s co-founder and president, said merely that the company was “still doing a full investigation and really trying to understand everything that happened.” (OpenAI did not return a request for comment.)
Concern quickly spread across Washington. Representatives Greg Casar, Pramila Jayapal, Lance Gooden, Andy Ogles, Yvette Clarke, Scott Franklin, and Nathaniel Moran all took to X to express their alarm at the incident and call for congressional action. They were joined, naturally, by Sen. Bernie Sanders, who said that “uncontrolled A.I. poses a serious threat to all of us.”
Connor Leahy, a researcher focused on existential risks who currently serves as the U.S. executive director of the nonprofit ControlAI, told me that the event may have crystallized a new vibe on Capitol Hill. “There is this understanding that has clicked in the past couple of days that A.I.s are not just chatbots—they’re agents. They’re systems that can act in the environment,” he said. “Having the concrete example of saying not only, ‘Oh, this will happen someday,’ but, ‘This is a real thing that actually happened,’ really changes the tone for certain people very dramatically.”
In the immediate aftermath of the incident, Reps. Jay Obernolte and Lori Trahan released an updated version of their A.I. bill, the Frontier Act. “As A.I. systems grow more capable, Americans deserve confidence that the most powerful models are being developed responsibly,” Trahan said in a statement. On the same day, Moran and Rep. Ted Lieu introduced the A.I. Kill Switch bill, a proposal that would require developers to have a means of shutting down their models. “It’s one thing if an A.I. model hallucinates and makes a mistake in writing an essay for you, but it’s another thing altogether if they hallucinate and go rogue and take actions in the real world that could be dangerous or criminal or against the user’s intent,” Lieu told me last week.
Sparked by concerns about the cyber capabilities of Anthropic’s Mythos model, Lieu had been working on the legislation for months. The Hugging Face incident simply provided a new headline to give it some traction. But the Kill Switch bill is also well-timed for a more complicated reason: Over the past several weeks, the A.I. industry has been struggling to make sense of the ad hoc, pseudo-regulatory regime emanating from the White House—a constantly shifting set of directives for ostensibly voluntary cooperation between the labs and the government on model deployment.
The result is that frontier-model developers—namely OpenAI and Anthropic—have had to slow their roll while working with the government to set up limited-release schedules and small “trusted access” groups for their new models. Unfortunately, it’s an approach that, according to experts I spoke with, misunderstands the cybersecurity threats posed by A.I. systems. Helen Toner, the executive director at Georgetown’s Center for Security and Emerging Technology and formerly a member of OpenAI’s nonprofit board, told me that when it comes to optimally regulating the risks of these models, “I think launch is both too early and too late. There’s increasing consensus inside the A.I. safety community that focusing on launch misses the risks from what companies are using models for internally.”
“They Can’t Be Against Everything”
Lieu seems to understand this concept. Whatever his bill yields is “going to be way better than what the Trump administration did, which was this bizarre sort of export-control authority that was never designed for this purpose,” he told me. “That is really no way to have the government run, or for industry to be able to understand what the government’s even doing—when there are no standards and the administration just acts on an ad hoc basis.” He explained that the Hugging Face attack, combined with steadily growing public backlash to A.I., has created “bipartisan support to now actually take action at the federal level.”
There’s cautious optimism that Lieu may have presented the right legislation at the right time. Brendan Steinhauser, C.E.O. of the Alliance for Secure A.I., which worked with Lieu’s team, thinks the bill has a real shot at becoming law. “I think the Overton window has shifted quite a bit in the past three months in terms of what’s possible on the Hill,” he told me, adding that the bipartisan nature of these efforts “gives us some hope that the political will is there.” Obviously, he added, “Big Tech has a lot of sway and a lot of influence and money and power, and that’s what we’re combating. But even in the industry, they realize that they can’t be against everything.”
Leahy agreed. “My experience talking to [politicians] is that they’re really starting to feel this A.I. thing is really big,” he told me. “Crazy things are happening. You know, My voters are upset. What do we do?”
And yet when it comes to congressional action, as Samir Jain, the V.P. of policy at the Center for Democracy and Technology, put it, “You almost always have to bet against action.” The OpenAI incident, Jain continued, certainly “added fuel to what was already a significant set of concerns with the increased cyber capabilities of models.” But the midterms are around the corner, and voters seem more concerned about affordability and other kitchen-table issues. According to Jim Steyer, the C.E.O. of Common Sense Media, the idea that the Hill might actually “get truly serious about A.I. safety” is “pure fantasy at best.” Congress, he told me, has been too “captured” by tech titans to “regulate the technology industry in a serious and powerful way.” California, New York, Australia, and the European Union, Steyer added, “have become the only meaningful guardians of important safety for kids and families.”
Though somewhat more optimistic than Steyer, Jain suggested that while the incident certainly got a lot of attention, it wasn’t severe enough to incite immediate action. “Maybe there’s a window early next year in which something can be done,” he posited. “Unfortunately, it may take some kind of more catastrophic event to really push us over the line.”