WEBVTT - Surveying the AI Threat Landscape

0:00:00.120 --> 0:00:03.960
<v Speaker 1>This is Bloomberg Business Week with Carol Messer and Tim

0:00:04.000 --> 0:00:06.240
<v Speaker 1>Stenebeck on Bloomberg Radio.

0:00:06.720 --> 0:00:09.159
<v Speaker 2>Jennifer, I don't know if you saw this making the

0:00:09.200 --> 0:00:11.799
<v Speaker 2>rounds on Twitter last week, but this video went viral.

0:00:11.840 --> 0:00:13.280
<v Speaker 1>It racked up more than a million views.

0:00:13.800 --> 0:00:16.640
<v Speaker 2>It purports to be from a company that calls itself

0:00:16.720 --> 0:00:20.200
<v Speaker 2>quote the world's first AI powered news network. Now, I

0:00:20.239 --> 0:00:23.560
<v Speaker 2>wasn't able to independently confirm that this video is indeed

0:00:23.720 --> 0:00:26.160
<v Speaker 2>all AI, but it's still worth taking a listen to

0:00:26.680 --> 0:00:29.160
<v Speaker 2>kind of see where we are in this day and age.

0:00:29.240 --> 0:00:29.760
<v Speaker 1>Check this out.

0:00:31.600 --> 0:00:36.440
<v Speaker 3>Hello, and welcome to Channel one, a new way of consuming, reporting,

0:00:36.640 --> 0:00:40.680
<v Speaker 3>and thinking about the news powered by artificial intelligence, all

0:00:40.720 --> 0:00:43.760
<v Speaker 3>presented by our team of AI generated reporters.

0:00:43.880 --> 0:00:46.919
<v Speaker 4>Will show you how AI powers and empowers our newsroom

0:00:47.200 --> 0:00:52.280
<v Speaker 4>to deliver journalism that's fast, trustworthy, and accurate. Let's start

0:00:52.320 --> 0:00:55.320
<v Speaker 4>with our reporters. You can hear us and see our

0:00:55.360 --> 0:00:58.639
<v Speaker 4>lips moving, but no one was recorded saying what we're

0:00:58.640 --> 0:01:02.440
<v Speaker 4>all saying, powered by sophisticated systems behind the scenes.

0:01:03.280 --> 0:01:06.400
<v Speaker 2>Okay, so again, I was not able to independently confirm

0:01:06.400 --> 0:01:08.600
<v Speaker 2>that this is indeed all AI, because if you were

0:01:08.640 --> 0:01:11.800
<v Speaker 2>watching that it looks like actual people, but there are

0:01:11.800 --> 0:01:14.920
<v Speaker 2>some clips of people of the actual anchors, like very

0:01:14.959 --> 0:01:16.720
<v Speaker 2>quickly speaking different languages.

0:01:17.240 --> 0:01:18.440
<v Speaker 1>It's all pretty wild.

0:01:18.760 --> 0:01:21.160
<v Speaker 2>So depending on your view, it could either be really

0:01:21.200 --> 0:01:24.280
<v Speaker 2>exciting or if you react like me and then everybody

0:01:24.319 --> 0:01:26.640
<v Speaker 2>I showed this video too in the newsroom, it's pretty

0:01:26.720 --> 0:01:28.800
<v Speaker 2>scary stuff. Let's see what our next guest has to

0:01:28.800 --> 0:01:30.520
<v Speaker 2>say about all this, because he's got more than twenty

0:01:30.560 --> 0:01:32.560
<v Speaker 2>five years of experience on the front lines when it

0:01:32.560 --> 0:01:36.520
<v Speaker 2>comes to cybersecurity, which in this day and age increasingly

0:01:36.560 --> 0:01:40.360
<v Speaker 2>looks at threats posed by so called deep fakes, AI

0:01:40.440 --> 0:01:45.959
<v Speaker 2>and misinformation. We've got Chester Wiznski, Global Field CTO at Sophos.

0:01:46.319 --> 0:01:48.280
<v Speaker 2>He joins us on Zoom from Vancouver.

0:01:48.400 --> 0:01:50.640
<v Speaker 1>Chester. Good to have you with us this afternoon. How

0:01:50.680 --> 0:01:53.560
<v Speaker 1>are you. I'm well, good to be here.

0:01:53.560 --> 0:01:56.480
<v Speaker 5>It's getting close to the holiday times, so it's kind

0:01:56.480 --> 0:01:59.920
<v Speaker 5>of an exciting time to get a chance to have

0:02:00.040 --> 0:02:02.960
<v Speaker 5>a chat about all the things we're going to be

0:02:02.960 --> 0:02:05.160
<v Speaker 5>seeing in our social media feeds the next few weeks.

0:02:05.280 --> 0:02:05.480
<v Speaker 1>Yeah.

0:02:05.520 --> 0:02:07.800
<v Speaker 2>The reason I brought that up is because this is

0:02:07.840 --> 0:02:09.760
<v Speaker 2>something that I signed my social media feed over the

0:02:09.840 --> 0:02:12.079
<v Speaker 2>last couple of weeks, and again I reached out to

0:02:12.120 --> 0:02:14.040
<v Speaker 2>this company, and I wasn't able to confirm whether or

0:02:14.080 --> 0:02:17.640
<v Speaker 2>not it was all AI, but you got to see

0:02:17.639 --> 0:02:19.919
<v Speaker 2>some of it, You got to listen to some of it.

0:02:19.960 --> 0:02:22.880
<v Speaker 2>Is that the type of deep fake that we're seeing

0:02:22.919 --> 0:02:25.680
<v Speaker 2>these days, Well.

0:02:25.480 --> 0:02:28.639
<v Speaker 5>I think that's on the sophisticated end of the spectrum.

0:02:28.680 --> 0:02:31.640
<v Speaker 5>And it's important to remember that wasn't probably created in

0:02:31.720 --> 0:02:34.079
<v Speaker 5>real time. You know, it wasn't something that they could

0:02:34.160 --> 0:02:36.560
<v Speaker 5>just whip up in thirty seconds, right, I mean there was.

0:02:37.000 --> 0:02:40.400
<v Speaker 1>Clearly some effort very high that. Yeah. Yeah.

0:02:40.440 --> 0:02:42.320
<v Speaker 5>But I mean, on the other hand, like this is

0:02:42.360 --> 0:02:44.000
<v Speaker 5>how far this has come, right, if you have the

0:02:44.080 --> 0:02:47.959
<v Speaker 5>resources to want to create a realistic audio and video,

0:02:48.080 --> 0:02:50.760
<v Speaker 5>and certainly, as we've all seen with chat GPT in

0:02:50.800 --> 0:02:54.880
<v Speaker 5>the last year text, if you've got the resources, it

0:02:55.240 --> 0:02:58.000
<v Speaker 5>is getting really close to the point where it's indiscernible

0:02:58.000 --> 0:03:00.960
<v Speaker 5>from real content. I suspect that that entire video is

0:03:01.000 --> 0:03:03.320
<v Speaker 5>all done an AI. While you were playing the role

0:03:03.320 --> 0:03:05.680
<v Speaker 5>in there, I was hearing some telltale things in the

0:03:05.720 --> 0:03:07.880
<v Speaker 5>audio that my ear kind of knows to listen for.

0:03:08.000 --> 0:03:10.840
<v Speaker 5>But it's pretty convincing, isn't it.

0:03:11.160 --> 0:03:13.760
<v Speaker 6>What were some of the telltale things to listen for

0:03:13.880 --> 0:03:16.760
<v Speaker 6>because I must confess they passed me by completely.

0:03:18.080 --> 0:03:20.240
<v Speaker 5>Yeah, well, it's getting better all the time, certainly when

0:03:20.240 --> 0:03:24.680
<v Speaker 5>we're looking at AI generated photos and even video. Sometimes

0:03:24.840 --> 0:03:29.280
<v Speaker 5>AI historically has had a pretty bad approach to creating

0:03:29.320 --> 0:03:32.040
<v Speaker 5>straight lines. You'd think straight lines are the simplest thing

0:03:32.080 --> 0:03:34.320
<v Speaker 5>for a computer to draw, ever, but the way AI

0:03:34.480 --> 0:03:36.360
<v Speaker 5>is trained, it's actually really hard to teach it to

0:03:36.400 --> 0:03:39.400
<v Speaker 5>make straight lines. And with human faces, it often makes

0:03:39.440 --> 0:03:42.240
<v Speaker 5>mistakes like one ear lobe being attached and the other

0:03:42.280 --> 0:03:44.600
<v Speaker 5>one being detached, or too many teeth in the mouth,

0:03:44.760 --> 0:03:47.360
<v Speaker 5>or you know, there's some telltale things that frequently messes

0:03:47.440 --> 0:03:50.800
<v Speaker 5>up too many fingers on a hand, for example. But

0:03:50.960 --> 0:03:52.440
<v Speaker 5>these are all getting better all the time. And in

0:03:52.480 --> 0:03:54.760
<v Speaker 5>the audio the things I was listening for, there was

0:03:55.040 --> 0:03:57.720
<v Speaker 5>there was there's a crispness to each word in the

0:03:57.760 --> 0:04:00.360
<v Speaker 5>way that they're separated, and it almost had a bit

0:04:00.360 --> 0:04:04.880
<v Speaker 5>of a slightly robotic tone when it's trying to inflect emotion.

0:04:05.040 --> 0:04:07.280
<v Speaker 5>That is just feels a little bit off.

0:04:07.160 --> 0:04:09.000
<v Speaker 1>When I'm listening to it. Right when I heard the.

0:04:09.000 --> 0:04:12.320
<v Speaker 5>Anchor saying regular words, it sounded pretty plausible. But then

0:04:12.360 --> 0:04:14.600
<v Speaker 5>if I listened really carefully when she was trying to

0:04:14.640 --> 0:04:17.600
<v Speaker 5>inflect some emotion. It had a bit of a robotic

0:04:17.640 --> 0:04:18.600
<v Speaker 5>on natural feel to.

0:04:18.560 --> 0:04:19.600
<v Speaker 1>It, still so ched.

0:04:19.640 --> 0:04:22.200
<v Speaker 2>It all raises the question how this stuff can be

0:04:22.320 --> 0:04:25.159
<v Speaker 2>used for malicious purposes moving forward. I mean my first

0:04:25.200 --> 0:04:29.640
<v Speaker 2>thought is misinformation and disinformation, the idea that videos can

0:04:29.680 --> 0:04:32.720
<v Speaker 2>be created that you can just not trust, whether that's

0:04:32.760 --> 0:04:36.440
<v Speaker 2>of a world leader saying something or doing something, or

0:04:36.480 --> 0:04:39.360
<v Speaker 2>of a celebrity saying something that just isn't real.

0:04:41.000 --> 0:04:44.200
<v Speaker 5>Yeah, And I think that's really the real problem here

0:04:44.279 --> 0:04:48.080
<v Speaker 5>is this allows this content to be created at a

0:04:48.120 --> 0:04:50.520
<v Speaker 5>scale that it hasn't been able to be done before. Right,

0:04:50.560 --> 0:04:53.760
<v Speaker 5>if you and I have been speaking last year at Christmas,

0:04:53.880 --> 0:04:56.599
<v Speaker 5>or even the year before at Christmas, it would have

0:04:56.600 --> 0:04:59.400
<v Speaker 5>been pretty unlikely that a commercial company would have been

0:04:59.440 --> 0:05:01.719
<v Speaker 5>able to produce that video that you played a little

0:05:01.760 --> 0:05:05.440
<v Speaker 5>clip of. And it's accelerating so fast now and it's

0:05:05.440 --> 0:05:07.560
<v Speaker 5>getting so hard to tell the good from the bad.

0:05:07.560 --> 0:05:10.120
<v Speaker 5>It also means it can be done at a volume

0:05:10.279 --> 0:05:13.680
<v Speaker 5>that humans would be unable of creating. Right When we

0:05:13.720 --> 0:05:18.240
<v Speaker 5>think about a video like that, that's something that now

0:05:18.279 --> 0:05:21.880
<v Speaker 5>can be done in hours and with's a modest amount

0:05:21.920 --> 0:05:25.480
<v Speaker 5>of money, whereas before, you know, it would have required

0:05:25.480 --> 0:05:28.320
<v Speaker 5>a ridiculous amount of resources to even attempt to generate

0:05:28.560 --> 0:05:32.680
<v Speaker 5>something that believable. So while it's not something you and

0:05:32.680 --> 0:05:34.279
<v Speaker 5>I are going to be able to create that newsroom

0:05:34.279 --> 0:05:37.000
<v Speaker 5>clip tomorrow, maybe if we talk again a year from now,

0:05:37.080 --> 0:05:38.000
<v Speaker 5>we might be able.

0:05:37.800 --> 0:05:39.480
<v Speaker 1>To even be able to do that as individuals.

0:05:39.480 --> 0:05:43.159
<v Speaker 5>So this is going to be very challenging moving forward

0:05:43.200 --> 0:05:46.560
<v Speaker 5>because it's not something that's easily to disprove. So I

0:05:46.600 --> 0:05:49.200
<v Speaker 5>think we need to be working with our media sources

0:05:49.560 --> 0:05:52.880
<v Speaker 5>toward making sure the real content can be proven to

0:05:52.920 --> 0:05:57.440
<v Speaker 5>be real as a way of, you know, allowing us

0:05:57.480 --> 0:06:00.240
<v Speaker 5>to more easily discern what's real what isn't you know what?

0:06:00.279 --> 0:06:02.599
<v Speaker 6>This is actually leading up to my next question because,

0:06:02.680 --> 0:06:04.760
<v Speaker 6>as you say, a year from now, who knows what

0:06:04.800 --> 0:06:07.000
<v Speaker 6>the capabilities are going to be, but from what you're saying,

0:06:07.320 --> 0:06:09.920
<v Speaker 6>they're going to be much better. And then if we're

0:06:09.920 --> 0:06:12.960
<v Speaker 6>looking at next Christmas when we have this conversation, the

0:06:13.040 --> 0:06:16.440
<v Speaker 6>thing that will have just happened before next Christmas is

0:06:16.480 --> 0:06:21.480
<v Speaker 6>the US election, and it's already very contentious, and we've

0:06:21.480 --> 0:06:26.040
<v Speaker 6>got to ask what can the average voter, you know,

0:06:26.520 --> 0:06:30.760
<v Speaker 6>US journalist, you know, policymakers, how are we supposed to

0:06:30.760 --> 0:06:33.400
<v Speaker 6>be able to tell what's real and what's fake as

0:06:33.400 --> 0:06:37.200
<v Speaker 6>the election approaches, and what are the risks to a

0:06:37.240 --> 0:06:40.280
<v Speaker 6>free and fair election if we're not able to really

0:06:40.320 --> 0:06:42.000
<v Speaker 6>get on top of this problem straight away.

0:06:43.600 --> 0:06:45.000
<v Speaker 5>Well, I think we just need to get in a

0:06:45.000 --> 0:06:48.960
<v Speaker 5>better habit of going to trusted, verifiable sources, whether that

0:06:49.760 --> 0:06:52.560
<v Speaker 5>if it's about a candidate or an election campaign, then

0:06:52.600 --> 0:06:55.919
<v Speaker 5>making sure that we're getting the information from the campaign itself,

0:06:56.000 --> 0:06:57.880
<v Speaker 5>so that we know it's real from that candidate. And

0:06:58.080 --> 0:07:00.719
<v Speaker 5>you know, we've been seeing on television ads for years

0:07:00.720 --> 0:07:05.400
<v Speaker 5>that this ad was approved by whoever the candidate is, right,

0:07:05.800 --> 0:07:09.039
<v Speaker 5>there is an official channel for receiving that information so

0:07:09.080 --> 0:07:11.760
<v Speaker 5>that we can verify that it's definitely from a campaign

0:07:11.840 --> 0:07:14.800
<v Speaker 5>or a party or that kind of thing. And of

0:07:14.800 --> 0:07:17.320
<v Speaker 5>course our new sources, I mean Bloomberg is a great

0:07:17.320 --> 0:07:20.240
<v Speaker 5>example of that. But I mean, no matter what your

0:07:20.280 --> 0:07:23.600
<v Speaker 5>politics are, that there's a lot of trusted journalism out there,

0:07:24.240 --> 0:07:27.280
<v Speaker 5>and going to those sources to verify a story before

0:07:27.320 --> 0:07:29.520
<v Speaker 5>we repost things on social media I think is going

0:07:29.560 --> 0:07:33.360
<v Speaker 5>to be a really important step because trusting random accounts

0:07:33.360 --> 0:07:37.000
<v Speaker 5>on x and Meta and TikTok is probably going to

0:07:37.040 --> 0:07:38.200
<v Speaker 5>lead us in a dangerous direction.

0:07:38.920 --> 0:07:42.040
<v Speaker 2>Yeah, that is certainly true, and I'm not optimistic about

0:07:42.080 --> 0:07:47.360
<v Speaker 2>the way that these platforms have allowed misinformation to flourish.

0:07:47.560 --> 0:07:52.360
<v Speaker 2>Hey chat, I promised i'd ask you, as somebody who's

0:07:52.360 --> 0:07:54.800
<v Speaker 2>been on the front lines of cybersecurity for twenty five years,

0:07:55.680 --> 0:07:57.720
<v Speaker 2>what worries you, what keeps you up at night?

0:08:00.120 --> 0:08:03.480
<v Speaker 5>Well, the quality of this stuff and the quantity of

0:08:03.600 --> 0:08:07.000
<v Speaker 5>that can be created is going to really blur the

0:08:07.040 --> 0:08:09.320
<v Speaker 5>lines for a lot of people between you know, the

0:08:09.360 --> 0:08:12.480
<v Speaker 5>reality and this misinformation, and a lot of us have

0:08:12.560 --> 0:08:16.040
<v Speaker 5>gotten too comfortable. You know, we don't specifically study people

0:08:16.080 --> 0:08:19.600
<v Speaker 5>trying to manipulate elections. We're worried about people being manipulated

0:08:19.640 --> 0:08:22.480
<v Speaker 5>into scams, getting you know, their money stolen from them,

0:08:22.680 --> 0:08:25.760
<v Speaker 5>getting their computers infected with viruses, that kind of thing.

0:08:25.840 --> 0:08:26.440
<v Speaker 1>Right, And when we.

0:08:26.400 --> 0:08:30.240
<v Speaker 5>Look at the things that people used to pay attention to,

0:08:30.360 --> 0:08:32.080
<v Speaker 5>you know, they read their email and they look for

0:08:32.120 --> 0:08:34.400
<v Speaker 5>the spelling mistakes. So this email doesn't really look like

0:08:34.440 --> 0:08:37.400
<v Speaker 5>it's from Bank of America. All those tell tale signs

0:08:37.400 --> 0:08:40.360
<v Speaker 5>are out the window now, right, There's really no discernible

0:08:40.400 --> 0:08:44.719
<v Speaker 5>difference between computer generated content that a criminal can make

0:08:44.840 --> 0:08:47.719
<v Speaker 5>tens of thousands of inn minutes and the real thing

0:08:47.720 --> 0:08:51.160
<v Speaker 5>that you're really getting from your financial institution or anybody

0:08:51.160 --> 0:08:54.559
<v Speaker 5>else that you do commerce with. So I think us

0:08:54.720 --> 0:09:00.320
<v Speaker 5>educating society on to be more suspicious and to double

0:09:00.400 --> 0:09:03.160
<v Speaker 5>check things before they believe them is going to be

0:09:03.160 --> 0:09:04.880
<v Speaker 5>important part of it. And more than that, I think

0:09:04.880 --> 0:09:06.720
<v Speaker 5>a lot of it comes back on the industry itself,

0:09:06.760 --> 0:09:09.160
<v Speaker 5>on the AI industry, and on us people in security

0:09:09.720 --> 0:09:12.840
<v Speaker 5>to build better tools to sort through this stuff so

0:09:12.880 --> 0:09:16.480
<v Speaker 5>that the majority of what ends up in your inbox

0:09:16.480 --> 0:09:17.120
<v Speaker 5>is signal, not.

0:09:17.160 --> 0:09:20.120
<v Speaker 6>Noise, just real quick thirty seconds. What do you think

0:09:20.280 --> 0:09:23.240
<v Speaker 6>regulators and social media companies should be doing about this?

0:09:24.760 --> 0:09:25.600
<v Speaker 1>Yeah, that's challenging.

0:09:25.600 --> 0:09:27.959
<v Speaker 5>I mean, the Biden White House has published in an

0:09:27.960 --> 0:09:30.880
<v Speaker 5>executive order at the end of October that covers a

0:09:30.880 --> 0:09:34.559
<v Speaker 5>lot of AI usage and is moving toward marking AI

0:09:34.640 --> 0:09:37.400
<v Speaker 5>generated content with a watermark. Certainly when it's produced by

0:09:37.400 --> 0:09:41.439
<v Speaker 5>the federal government. That's encouraging. But of course, what I'm

0:09:41.480 --> 0:09:43.200
<v Speaker 5>really worried about is people who aren't going.

0:09:43.120 --> 0:09:43.800
<v Speaker 4>To follow the law.

0:09:43.960 --> 0:09:46.920
<v Speaker 5>Right if we're worried about Russians and influencing our next election,

0:09:47.520 --> 0:09:49.679
<v Speaker 5>they're not going to abide by an executive order from

0:09:49.679 --> 0:09:51.760
<v Speaker 5>the White House to be good boys and good girls.

0:09:52.520 --> 0:09:54.280
<v Speaker 5>That's just not going to fly right. We have to

0:09:54.320 --> 0:09:58.400
<v Speaker 5>be able to, I think, invest our effort in proving

0:09:58.480 --> 0:10:02.440
<v Speaker 5>authenticity rather than trying to prove an inauthentic content.

0:10:02.679 --> 0:10:05.199
<v Speaker 2>Chester really appreciate you taking the time. That's Chet Wiznsky,

0:10:05.280 --> 0:10:08.040
<v Speaker 2>Global Field CTO at Sophos joining us from Vancouver.