Latest / DEEP INSIGHT / Ex-Pentagon AI chief drops bombshell system autonomously hacked outer target

Transcript

Ex-Pentagon AI chief drops bombshell system autonomously hacked outer target · DEEP INSIGHT

• 0:04
  Speaker 1: Let's talk about Anthropic. They did something that was.
• 0:09
  Speaker 3: Very bad, three months ago. And we shut them down.
• 0:14
  Speaker 3: We shut them down fast. We have guardrails. The biggest
• 0:17
  Speaker 3: guardrails are having people that are just as smart, because
• 0:20
  Speaker 3: nobody understands it. We're making strides in every business right
• 0:23
  Speaker 3: now that would take, in many cases, 100 years to
• 0:26
  Speaker 3: get to. The kind of things that are coming out
• 0:29
  Speaker 3: are incredible. Now, we want guardrails, but we— it' s a very
• 0:32
  Speaker 3: fine line. You know, China doesn't have guardrails, by the way.
• 0:36
  Speaker 3: You know, the guardrail is President Xi. That's the guardrail.
• 0:40
  Speaker 4: President Trump says AI needs guardrails, but not at the
• 0:43
  Speaker 4: expense of American innovation. Now the CEO of Anthropic is
• 0:47
  Speaker 4: calling on AI giants to slow down, warning that technology
• 0:51
  Speaker 4: could outpace our.
• 0:52
  Speaker 5: Ability to control it.
• 0:54
  Speaker 4: In an alarming essay, the anthropic leader actually wrote this, quote,
• 0:59
  Speaker 4: not building the technology deprives humanity of benefits or simply
• 1:03
  Speaker 4: places AI in the hands of authoritarian powers. While building
• 1:06
  Speaker 4: it too fast is reckless, it's my worry that in
• 1:10
  Speaker 4: six to 12 months, such a swarm could be capable
• 1:13
  Speaker 4: of taking over the entire Internet and that the scale
• 1:17
  Speaker 4: of damage would continue to increase from there. And top
• 1:21
  Speaker 4: executives from major AI companies quickly got on board. OpenAI
• 1:25
  Speaker 4: CEO Sam Altman and Elon Musk both backing Anthropit's caution.
• 1:30
  Speaker 4: We're going to bring in Pentagon AI Director Mark Beal. Mark,
• 1:34
  Speaker 4: thanks for joining us this morning.
• 1:37
  Speaker 5: So let's get right down to it.
• 1:39
  Speaker 4: Is AI doing things that we are.
• 1:43
  Speaker 5: Not asking it to do?
• 1:44
  Speaker 4: Is it acting independently of what the humans.
• 1:48
  Speaker 5: Are telling it to do?
• 1:50
  Speaker 1: Good morning. Great to be here. Unfortunately, that answer right
• 1:53
  Speaker 1: now is yes.
• 1:55
  Speaker 2: There was an incident called the Open AI Hugging Face Incident.
• 2:00
  Speaker 2: in which internal AI systems at this company, OpenAI, were
• 2:04
  Speaker 2: given a series of tests.
• 2:05
  Speaker 1: And those tests were really hard.
• 2:07
  Speaker 2: And what they decided to do instead of trying to
• 2:09
  Speaker 2: answer those tests was to hack out and autonomously attack
• 2:12
  Speaker 2: another company, which those AI agents thought.
• 2:15
  Speaker 1: Had the answer.
• 2:16
  Speaker 2: So already we're starting to see some alarming behaviors of
• 2:18
  Speaker 2: these systems.
• 2:20
  Speaker 1: Behaving in ways that human beings did not intend.
• 2:23
  Speaker 2: And so folks like Elon Musk are kind of looking
• 2:25
  Speaker 2: around the corner and seeing a pathway by which these
• 2:28
  Speaker 2: systems become vastly superhuman. And if we can't control them
• 2:32
  Speaker 2: under today's level of capability, you can see why there
• 2:35
  Speaker 2: could be some concern about tomorrows.
• 2:37
  Speaker 4: OK, I want to talk about regulation and what that
• 2:40
  Speaker 4: balance is. But before we do that, I want to
• 2:42
  Speaker 4: ask you one very specific question. Can AI tether itself
• 2:47
  Speaker 4: to weapon systems without us knowing? And if so, why
• 2:52
  Speaker 4: was there not a guardrail on that? That seems like
• 2:54
  Speaker 4: a very obvious thing.
• 2:57
  Speaker 2: I think theoretically, you have an intelligent system and it's very,
• 3:01
  Speaker 2: very capable.
• 3:03
  Speaker 1: There's a lot that it can do.
• 3:04
  Speaker 2: And I know this starts to sound a lot like
• 3:07
  Speaker 2: science fiction, but I think at this.
• 3:09
  Speaker 1: Stage, the science fiction movie is about here.
• 3:13
  Speaker 2: And so, again, when we're trying to figure out how
• 3:16
  Speaker 2: to put guardrails on these systems, and they're vastly smarter
• 3:19
  Speaker 2: than we are, it gets a little bit tricky. And
• 3:21
  Speaker 2: this is exactly what happened with that OpenAI Hugging Face
• 3:24
  Speaker 2: incident I referred to. You know, some of the top
• 3:26
  Speaker 2: cybersecurity engineers in the world are at OpenAI, and those
• 3:29
  Speaker 2: are the ones that designed cage for those models. And yet,
• 3:32
  Speaker 2: despite how smart those humans were, those AIs figured out
• 3:35
  Speaker 2: novel ways to.
• 3:36
  Speaker 1: Escape, to basically hack their way out.
• 3:39
  Speaker 2: And this is kind of the thing that's resulted in
• 3:42
  Speaker 2: some big questions around these things called guardrails, the ways
• 3:46
  Speaker 2: that you're going to try to keep people safe. And
• 3:48
  Speaker 2: I think it's also important to remember, like, we can
• 3:50
  Speaker 2: continue to diffuse AI in the economy and derive tremendous
• 3:53
  Speaker 2: amounts of benefit because there's, you know, prosperity for all Americans,
• 3:57
  Speaker 2: I think, is a thing that we're definitely seeing for
• 4:01
  Speaker 2: business productivity. At the same time, building the next more
• 4:05
  Speaker 2: powerful system is something that folks are starting to step
• 4:08
  Speaker 2: back and say, wait a minute, do we really want
• 4:09
  Speaker 2: to do this at the pace that we're doing it?
• 4:11
  Speaker 2: We need to make sure it's safe because ultimately public
• 4:13
  Speaker 2: safety is critically important.
• 4:16
  Speaker 5: Yeah, you don't want to throttle innovation.
• 4:18
  Speaker 4: But at the same time, I just wonder if the
• 4:20
  Speaker 4: people who are running this sometimes have godlike fantasies and
• 4:24
  Speaker 4: that they've gotten out of control. But let's look at
• 4:27
  Speaker 4: some of these headlines because they really speak to the
• 4:30
  Speaker 4: concern that some of these companies actually have. Here's NBC— on NBC,
• 4:36
  Speaker 4: government-linked accounts tried to use Claude for work that could
• 4:39
  Speaker 4: lead to bioweapons.
• 4:40
  Speaker 5: We just talked about that.
• 4:42
  Speaker 4: The Houthis used anthropic AI to try to build ballistic missiles.
• 4:46
  Speaker 4: And then Reuters reporting that an open AI agent attacked
• 4:49
  Speaker 4: Ruby Gems before hugging face incident.
• 4:53
  Speaker 5: That's what the researchers say.
• 4:55
  Speaker 4: So, we— Well, here's another— here' s some more headlines. Virginia County
• 5:01
  Speaker 4: expects— these are.
• 5:02
  Speaker 5: Some positive things, right?
• 5:03
  Speaker 4: So people are going, oh, there's a lot of money
• 5:04
  Speaker 4: that can be made here. Virginia County brought in $ 1. 3
• 5:08
  Speaker 4: billion in tax revenue from its data centers. You have
• 5:12
  Speaker 4: The Guardian saying that superhuman AI tools spotted a heart
• 5:16
  Speaker 4: disease in less than two seconds. And then small businesses
• 5:21
  Speaker 4: that adopt AI.
• 5:23
  Speaker 5: Are hiring more workers.
• 5:25
  Speaker 4: And yet, at the same time, you go to a
• 5:29
  Speaker 4: graduation ceremony and you see that there are, you know,
• 5:33
  Speaker 4: kids who are booing AI when it's brought up in
• 5:36
  Speaker 4: the commencement address. So what do those kids know that
• 5:39
  Speaker 4: we don't know about the future?
• 5:43
  Speaker 1: Intelligence, in some respects, is the ultimate dual-use capability.
• 5:48
  Speaker 2: And those headlines just sort of show exactly what we're
• 5:51
  Speaker 2: talking about here. Curing heart disease, curing many types of cancers,
• 5:58
  Speaker 2: business productivity, the fact that workers are more productive with
• 6:02
  Speaker 2: these tools, that we're using them to grow faster and
• 6:06
  Speaker 2: do more.
• 6:06
  Speaker 1: Of course, these things are great.
• 6:09
  Speaker 5: Is there a chance that this could just cause....
• 6:13
  Speaker 4: A concentration of power, not just that the machines are
• 6:17
  Speaker 4: going rogue and doing that, which is really troubling in
• 6:20
  Speaker 4: and of itself, but are you concerned that in the
• 6:23
  Speaker 4: end this will cause such a concentration of power that
• 6:27
  Speaker 4: we will lose our democracy in many ways?
• 6:31
  Speaker 2: Well, you know, I hate to speculate and fearmonger about
• 6:35
  Speaker 2: the future, but I think we're Americans. We can do
• 6:38
  Speaker 2: anything we set our mind to. And so some of
• 6:40
  Speaker 2: these risks, like concentration of power, could manifest themselves quite easily.
• 6:45
  Speaker 2: And if you have one or two companies that are
• 6:47
  Speaker 2: responsible for running the entire economy— in the future.
• 6:50
  Speaker 1: Then you have some serious questions about that.
• 6:53
  Speaker 2: That's why things like open source models, models that are
• 6:56
  Speaker 2: not necessarily owned by one company, but that can be
• 6:59
  Speaker 2: accessed by everybody. These things are going to be very,
• 7:02
  Speaker 2: very important as we navigate this potential future. And then,
• 7:05
  Speaker 2: by the way, at the end of the day, people
• 7:06
  Speaker 2: are kind of scared.
• 7:08
  Speaker 5: Just a quick yes or no.
• 7:10
  Speaker 4: Do you think China will get on board to bring about, like,
• 7:14
  Speaker 4: join us in putting in guardrails that both nations will have.
• 7:19
  Speaker 5: To adhere to?
• 7:22
  Speaker 2: As long as America leads and we are in the
• 7:24
  Speaker 2: lead and they know it, then they are going to
• 7:27
  Speaker 2: cooperate with us because we're going to negotiate from a
• 7:29
  Speaker 2: position of strength.
• 7:30
  Speaker 1: Okay.
• 7:31
  Speaker 4: Thank you so much. Appreciate you joining us this morning.
• 7:34
  Speaker 1: Thank you.
• 7:35
  Speaker 5: Hey everyone, I'm Emily Campagno.
• 7:37
  Speaker 6: Catch me and my co-hosts Harris Faulkner and Kayleigh McEnany
• 7:40
  Speaker 6: on Outnumbered every weekday at 12 p.m. Eastern or set
• 7:44
  Speaker 6: your DVR. Also, don't forget to subscribe to the Fox
• 7:47
  Speaker 6: News YouTube page for daily highlights.