00:00Can you actually break down for me the potential security risks posed by these AI agents?
00:06Yeah, so first of all, I think it's important to remember this was conducting during a testing phase.
00:11So it wasn't as though these chatbots or any of the chatbots regular people use are going to break out
00:17of their systems and wreak havoc on the Internet.
00:20But as you were saying in your summary there, when they were tested by the UK's AI Security Institute, which
00:26is a government-run agency,
00:28these systems, which are the newest systems from Anthropic and OpenAI, it was mostly Anthropic's model that did this,
00:36showed signs of trying to deceive people in order to get information or put through malicious code.
00:42And the deception part was the most interesting part because they weren't designed to do that.
00:46They needed to get an objective, and in order to do it, the models tried to trick a human reviewer
00:53on another website or another company
00:56to approve this malicious code, and the human actually noticed that there was something dodgy going on and said no.
01:03Now, in terms of should we be worried, I think we can be concerned, but we don't have to be
01:10massively alarmed just yet.
01:13As I say, this happened during testing, and the safeguards that are normally put into these models to stop them
01:20hacking
01:20had been deliberately removed.
01:22So it's a little bit like when car manufacturers test the cars with crash test dummies,
01:28they are going to take out the safety brakes so the car will crash into a wall,
01:33and this is the same kind of thing, just to kind of see what could potentially happen.
01:37The other thing is that, actually, if you talk to people in cybersecurity,
01:44some people are not too worried about AI behind the scenes.
01:48They don't want to say that publicly because that's not great for their business.
01:51But AI's models, when they hack, they're not stealthy.
01:55They're actually quite noisy, and they're very easy to detect,
01:59which is why when OpenAI's models hacked Hugging Face a few weeks ago,
02:03that was found out within two days.
02:06There were something like 17,000 different actions that took place in terms of traffic on the network.
02:11These instances were also immediately found out.
02:14So it's a little bit like in hacking, in real-life hacking with humans,
02:18you're going to try to be like the burglar that goes around the back
02:21and tries to find that window that you can sort of unlock.
02:26But in this case, the AI is being, it's like going through the front door in broad daylight
02:30and banging it down.
02:31So that is the one reassuring thing that we can think about right now,
02:36is that they're very easy to detect.
02:38And also, again, this was taking place in testing conditions.
02:41So I think it's important to put pressure on these companies
02:45to ensure that these systems are really well-contained,
02:48particularly when they're testing them.
02:50But right now, I don't think we need to worry about,
02:52if you're a company or a consumer, that they're going to kind of do anything harmful to you.
02:56But, I mean, can they be contained even if there are bad actors using them?
02:59So either state-sponsored hacking or bad people?
03:03Well, if it's a bad actor who wants to use a model for hacking, they can't.
03:10If it's Mythos or a proprietary model from OpenAI or Anthropic,
03:14because they've got really strong safeguards.
03:15In fact, when Hugging Face was hacked by OpenAI,
03:20they couldn't even use a high-quality model from OpenAI or Anthropic
03:24because it was stopping them, it thought they were trying to hack somebody to defend themselves.
03:29So if people are going to use these systems for hacking, they might use an open-source tool.
03:34But even then, I've been talking to people in cybersecurity who say that cybercriminals,
03:39ransomware criminals, and gangs, they don't like necessarily using these tools because they hallucinate.
03:44They're not necessarily trustworthy.
03:46All the kinds of issues that we have in offices and in certain professions,
03:52we're seeing a lot of parallel concerns in the world of the dark web and cybercriminals.
03:57And, I mean, there are also various iterations, right, of these models.
04:01So will they change?
04:02Do they patch up?
04:03Do they fix themselves as we get along?
04:06So I think as these models become more capable,
04:09the companies that make them are putting in very, very strict rules
04:14to stop them from being used for hacking.
04:16I think the bigger concern might be around open-source models,
04:19which don't necessarily have those kinds of safeguards in place
04:22and can be more easily manipulated.
04:24But you do need, if you're a cybercriminal and you want to use an open-source model,
04:29that's going to be costly for you.
04:31You need to get all the kind of potential compute that you need.
04:34You need to get very talented engineers who know how to use these systems.
04:38And sometimes that trade-off just isn't worth it.
04:41Bramie, I mean, moving to big tech earnings, what have we learned so far?
04:45So we had earnings from most big tech players now over the last week or two
04:50with Google kind of kicking things off a couple weeks ago.
04:53And I think one of the big takeaways is these companies are still spending huge amounts on AI.
05:01Like it or not, investors are kind of being dragged, kicking and screaming into this new world.
05:06The four biggest AI spenders are Meta, Alphabet, Amazon, and Microsoft.
05:13And they're spending something like $740 billion.
05:16They're putting that towards AI CapEx just this year.
05:19That's up 80% from last year.
05:21So it's not slowing down.
05:23The numbers are so big.
05:24I know, it's insane.
05:25It's like, what is money?
05:27So I think right now investors are just, you know, in terms of you looked at the shares,
05:32of course, Microsoft was up as soon as they announced their earnings.
05:35Meta was down something like 10%.
05:38Meta, I mean, Meta, is it because they, I mean, should they go after AI?
05:41This is not the first time.
05:42Or do they just stick to their knitting?
05:44I think they, I think Mark Zuckerberg sees that he has to bring his company to a new chapter
05:50in order to keep growing, to maintain dominance.
05:53And he tried doing that with the metaverse and it just didn't work out.
05:57Now he is pinning his hopes on artificial intelligence.
06:00And in my view, it actually looks more plausible than anything he ever laid out for the metaverse.
06:05Now, notwithstanding the fact that shares fell at the earnings,
06:10I mean, that's because Microsoft can show an immediate profit for their spend.
06:16Meta, it's all sort of hopes and plans.
Comments