【情報漏洩】中国製AI「GLM」の悪用か?タイムズなど大手の被害続発・・・「最新AI」の脅威と自分の情報を守る「3つの防衛策」 — Transcript
Full transcript
- 0:00Hello, everyone. Today, we will take a
- 0:03deep dive into the cyberattacks
- 0:05targeting companies and the massive
- 0:08data breaches being reported daily.
- 0:10Starting with the data breach of up to
- 0:12approximately 6.6 million records at
- 0:14Times Car, announcements of damages
- 0:15targeting companies and organizations
- 0:17have been appearing one after another
- 0:19over the past few days. Joining us in
- 0:21the studio today are Mr. Sho Nakatani
- 0:23from AgenticSec, an expert in AI within
- 0:26the field of cybersecurity, and Ms. Umi
- 0:28of TV Tokyo, who covers the technology
- 0:30sector. It’s a pleasure to have you
- 0:33both with us.
- 0:33>> Thank you for having us.
- 0:35>> You appeared in our previous
- 0:36explanatory video for Claude Multus,
- 0:38didn't you? Yes. Mr. Nakatani, you run
- 0:41a security business, don't you?
- 0:43>> Yes. That's right. At AgenticSec, we
- 0:46provide products that help protect our
- 0:49customers 'cybersecurity by using the
- 0:52same techniques attackers use to stay
- 0:54one step ahead and identify risks early
- 0:57.
- 0:57>> Uh-huh. I see, so you are providing
- 1:00support to those companies.
- 1:02>> You're well-versed in the intersection
- 1:04of AI and security, I presume.
- 1:05>> That's right. Exactly. Since attackers
- 1:08nowadays are using AI, we offer
- 1:10products that simulate what happens
- 1:12when AI is used to carry out security
- 1:14attacks, so we can demonstrate those
- 1:17risks.
- 1:17>> I see. Today, I'd like you to explain
- 1:20how AI is being used behind the scenes.
- 1:23>> Certainly.
- 1:24>> Now, regarding the data breaches being
- 1:27reported day after day, their scale and
- 1:29scope continue to expand. At discount
- 1:31store operator MRIS Tarmarks, up to
- 1:34approximately 1.73 million records. At
- 1:37GMO's survey site, up to approximately
- 1:39950,000 records, and at Yakiniku King,
- 1:4210.78 million records, which accounts
- 1:45for nearly all of their members. It has
- 1:47been announced that 550,000 records at
- 1:50an Asahi subsidiary and 100,000 records
- 1:53at Citizen Watch may have been leaked.
- 1:55Furthermore, ransomware attacks on
- 1:57institutions like Daiwa Securities and
- 2:00Osaka University have caused immense
- 2:02damage across all industries. In
- 2:04response to this situation, Digital
- 2:06Minister Furukawa called on the public
- 2:07at a post-Cabinet meeting press
- 2:09conference to be conscious of
- 2:10protecting their own information, and
- 2:12the Chief Cabinet Secretary has also
- 2:14expressed that the government is taking
- 2:16this very seriously and heightening its
- 2:18vigilance. It seems like these
- 2:19incidents have surged over the past few
- 2:21weeks, but what is your take on this,
- 2:23Mr. Nakatani?
- 2:24>> Well, that’s true. I’m on X
- 2:26whenever I have a spare moment, and
- 2:28starting around the end of September,
- 2:31with the Times incident leading the way
- 2:33, we’ve seen news of data leaks
- 2:35almost every single day. As someone who
- 2:38runs a company dedicated to security,
- 2:40it truly pains me to see this happening
- 2:43. Yeah.
- 2:44>> Should we view this as a genuine
- 2:46increase over the last few weeks, or is
- 2:48it more a case of existing issues
- 2:49finally coming to light?
- 2:51>> I’d say it’s both. I believe the
- 2:53pattern of these issues coming to light
- 2:56is definitely a factor. We are seeing a
- 2:59significant rise in reported cases, but
- 3:02the timing of the actual attacks is a
- 3:04separate matter. Seeing so many other
- 3:08companies being targeted, some firms
- 3:10are conducting internal reviews,
- 3:12telling their teams to check again if
- 3:15they’ve been hit. I suspect that in
- 3:19some cases, they’ve discovered they
- 3:21were actually compromised months ago.
- 3:25So, while it is a fact that attacks are
- 3:27increasing, the rise in incidents we
- 3:30see now involves both fresh attacks and
- 3:32older ones that are only just now being
- 3:35uncovered.
- 3:36>> Yeah. It feels as though corporate
- 3:39vigilance is also intensifying.
- 3:42According to a report by the National
- 3:44Police Agency, the number of suspicious
- 3:47access attempts classified as the
- 3:49preparatory stage of a cyberattack
- 3:51reached 13,687 per day per IP address
- 3:54in the first half of this year. This is
- 3:57a 50%increase compared to the same
- 3:59period last year, marking an all-time
- 4:02high. Mr. Nakatani, personally, I
- 4:04suspect that the latest AI technology
- 4:07is driving this increase in
- 4:09cyberattacks. What is your actual view
- 4:12on that?
- 4:14>> I personally share the same view; as AI
- 4:16models have continued to advance, I
- 4:19definitely feel there is a correlation
- 4:21with the current rise in attacks.
- 4:24Mm-hmm.
- 4:25>> However, it’s not yet clear whether
- 4:28the increase in attacks we're seeing
- 4:30now is definitively due to these newer
- 4:32models. There isn't enough evidence
- 4:34circulating in the world to make a
- 4:36definitive judgment just yet.
- 4:38>> Mm-hmm.
- 4:41>> Well, speaking of whether or not
- 4:43state-of-the-art AI models are being
- 4:45used, what I’m focusing on right now
- 4:47are open-weight models. These include
- 4:51open-weight models often produced in
- 4:53places like China, and one that's been
- 4:56a hot topic in the security community
- 4:58lately is the GLM 5.3 model. Compared
- 5:03to something like Anthropic's Mythos
- 5:06released in April, this model has a
- 5:08comparable level of cybersecurity
- 5:10attack capability, yet it lacks the
- 5:13guardrails that US frontier models—
- 5:16like Mythos, Fable, or Astra GPT-6 Sol
- 5:18—have to prevent dangerous
- 5:20cyberattacks. Detailed reports from
- 5:26Anthropic suggest that the rise in
- 5:29cyberattacks may be linked to the
- 5:31emergence of models like GLM 5.3 that
- 5:34possess high capabilities without those
- 5:36built-in guardrails. I personally
- 5:39believe this is likely a contributing
- 5:41factor.
- 5:42>> Mm-hmm.
- 5:43>> Is GLM 5.3 really that powerful?
- 5:46>> It is incredibly impressive, yes. We
- 5:48saw reports about how cybersecurity
- 5:50capabilities had shifted and improved
- 5:52drastically with Mythos back in April,
- 5:54and here we are in October, so in just
- 5:57about six months, that level of power
- 5:59has become available in an open-weight
- 6:01format. The fact that something
- 6:03appeared that is easy for attackers to
- 6:06use because it lacks guardrails was
- 6:08something we anticipated, but it
- 6:10happened faster than expected. Combined
- 6:13with the rise in these attacks, I feel
- 6:16that while the impact is as large as we
- 6:18imagined, it is also proving to be even
- 6:21more concerning than I had anticipated.
- 6:24>> Mm-hmm. To add a bit of context,
- 6:27regarding the point about "open weights
- 6:29," you have models like the well-known
- 6:32ChatGPT or the Claude series from
- 6:34Anthropic—which we've mentioned—
- 6:36that are considered closed models, and
- 6:39then there is the open-source category,
- 6:41but on the other hand, "
- 6:43>> Open weights" sits somewhere in between
- 6:45. In other words, since the weights are
- 6:48something users can tinker with a bit
- 6:50—whereas changing the settings of
- 6:52ChatGPT itself is difficult—with open
- 6:55weights, users can tune the model to
- 6:57their own preferences, is that right?
- 7:00>> Yes, that’s correct. Well, with an
- 7:02open-weight model, you can obviously
- 7:04just download it to your own
- 7:06environment and use it on a high-spec
- 7:08machine without tuning it. Plus, by
- 7:10toggling the weights—the connection
- 7:12values between the model’s parameters
- 7:14that you mentioned—you can, for
- 7:15instance, adjust the alignment, which
- 7:17is the part that tells the model not to
- 7:19do anything dangerous from a security
- 7:21standpoint. So, it is technically
- 7:23possible to tamper with those sorts of
- 7:26guardrails. For GLM-4, it currently
- 7:30doesn’t seem like there’s much need
- 7:32to go through the trouble of modifying
- 7:34those weights, but in the future, if
- 7:36guardrails and alignment for
- 7:38open-weight models become too strict,
- 7:40it wouldn't be surprising to see people
- 7:43take advantage of the open weights to
- 7:45modify them and potentially misuse the
- 7:47model. So, the extent of the guardrails
- 7:51provided is essentially left up to the
- 7:54morality of the development company,
- 7:56isn't it? I believe it really comes
- 7:58down to their morals, what they are
- 8:00aiming to achieve with their
- 8:02development, and what kind of users
- 8:04they want to reach.
- 8:05>> Well, for example, regarding Zhipu AI,
- 8:08the Chinese company behind GLM-4, I
- 8:10don't know the exact intentions behind
- 8:13their release. However, compared to US
- 8:16companies like OpenAI or Anthropic,
- 8:19performance remains their top priority
- 8:21—their primary focus is on catching
- 8:24up to and surpassing the US. Meanwhile,
- 8:27US companies have already reached a
- 8:30point where, in terms of cybersecurity
- 8:32and mathematical capabilities—at
- 8:34least within the scope of what
- 8:36benchmarks can measure—they've gone
- 8:38quite far. Instead of pushing those
- 8:42further and competing over
- 8:43harder-to-understand benchmarks, I
- 8:46believe US firms are now shifting their
- 8:48focus toward safety and broadening
- 8:51their general capabilities beyond just
- 8:53cyber and math.
- 8:54>> Ah, I see. So it’s your impression
- 8:58that China is still heavily focused on
- 9:00growth in the security sector?
- 9:01>> I can't say for sure since I don't have
- 9:04exact information, but regarding the
- 9:06GLM-4 model, I definitely feel there's
- 9:08a level of performance growth that at
- 9:10least suggests security capabilities
- 9:13are being targeted.
- 9:14>> Looking at it from your perspective,
- 9:15Mr. Nakatani, it's only been a few
- 9:17months since the release of the
- 9:18previous model, right?
- 9:19>> Yes. Yes. So, was it somewhat expected
- 9:25that China would release a model like
- 9:28this GLM that rivals frontier models,
- 9:30especially one with open weights?
- 9:35>> That's correct. I did expect it. As I
- 9:38mentioned in a previous interview,
- 9:40there has consistently been a trend
- 9:43where open-weight models catch up in
- 9:45performance a few months behind US
- 9:47frontier models like those from OpenAI
- 9:50or Anthropic—it’s generally said to
- 9:52be a lag of about six to ten months. So
- 9:55, with the release of models like
- 9:57Claude 3, I think I mentioned somewhere
- 10:00that this wouldn't just be something
- 10:02Anthropic could do, that OpenAI would
- 10:04follow suit, and that open-weight
- 10:06models would catch up in about half a
- 10:09year or so. Honestly, I feel that's
- 10:11exactly what ended up happening.
- 10:13>> Yeah.
- 10:14>> I see.
- 10:15>> But what is the reason why we can't
- 10:17definitively conclude that this whole
- 10:20series of incidents was caused by
- 10:22things like GLM or other AI models?
- 10:24>> Well, let's see. I'll explain this
- 10:27while showing you a diagram.
- 10:31>> The key point in this diagram is that
- 10:33the victim system on the far right
- 10:36receives attack traffic from two
- 10:38sources and cannot distinguish between
- 10:40them. These two sources, as shown,
- 10:44represent an AI-driven cyberattack at
- 10:47the top and a conventional, non-AI
- 10:50cyberattack at the bottom. Saying that
- 10:52they are indistinguishable...is what
- 10:54this illustration shows.
- 10:56>> Yes.
- 10:57>> Furthermore, if AI attacks and
- 10:59conventional attacks are
- 11:01indistinguishable, it's even more
- 11:03impossible to tell which AI was used
- 11:05for an AI-based attack. Therefore, I'm
- 11:08saying it's honestly difficult for the
- 11:10victim side to identify which model was
- 11:13used. What exactly is a harness?
- 11:16>> A harness is...well, before talking
- 11:18about the harness, if you look at the
- 11:20AI model on the left, this is the basic
- 11:23behavior you experience when talking to
- 11:25something like ChatGPT—you ask a
- 11:27question in a language, and a language
- 11:30response comes back. This is
- 11:32essentially the AI model receiving
- 11:34tokens and outputting tokens. But if it
- 11:39only outputs language, it's hard to
- 11:42perform things like cyberattacks, or in
- 11:45some cases, generating diagrams or
- 11:48images, so then, what exactly is...
- 11:52>> When giving a model the ability to
- 11:55execute computer programs, we attach
- 11:57what's called a harness to it;
- 11:59basically, it's a tool or a layer that
- 12:01enables the execution of code or source
- 12:04code.
- 12:04>> Mm. Yeah.
- 12:07>> So, there’s currently a two-step
- 12:09process: the AI model generates
- 12:11instructions, and then this control
- 12:13infrastructure, or harness, executes
- 12:15them.
- 12:18>> And you don't even have to consciously
- 12:20think about passing the model's output
- 12:22to the harness anymore; if you use
- 12:24tools like the famous coding agents
- 12:27Codex or Claude Code, the user just
- 12:29sends an instruction, it automatically
- 12:31reaches the model, the model processes
- 12:34the tokens, and then it calls the
- 12:39harness as needed to perform the action
- 12:42the user wants—that's how AI works
- 12:44now.
- 12:45>> Mm.
- 12:46>> If guardrails are in place, does that
- 12:48mean the AI model simply won't output
- 12:50the code to do those things in the
- 12:52first place?
- 12:53>> It’s actually both; there are parts
- 12:55prevented on the model side and parts
- 12:57prevented on the harness side. The
- 13:00prevention on the model side uses a
- 13:03technique often called alignment; when
- 13:05the model spits out tokens, it has been
- 13:08fine-tuned—or "aligned"—so that
- 13:10it's embedded with instructions like "
- 13:12don't say things like this," making it
- 13:15so those tokens generally don't even
- 13:17come out. But since that alone is a bit
- 13:20too weak as a guarantee, you can also
- 13:22have the harness block things like a
- 13:24blacklist, ensuring certain actions are
- 13:26never taken, and that collection of
- 13:28controls is what we call "rails." Mm.
- 13:31Yeah.
- 13:32>> Now, as for this GLM, while it does
- 13:35have guardrails, they are, in a sense,
- 13:37very low.
- 13:38>> I think that’s the situation.
- 13:40>> I see. Mm.
- 13:42>> I’d like to ask one more thing:
- 13:43regarding attacks from the AI side, or
- 13:46attacks using tools, or attacks by
- 13:47humans—even if you can't distinguish
- 13:49between them, for example...
- 13:51>> It's easy to imagine that with AI, you
- 13:53could just keep sending attacks
- 13:55indefinitely, like 24/7, so wouldn't
- 13:57there be a difference in volume? Is
- 13:59that not the case?
- 14:00>> Yes. Compared to old-school attacks
- 14:03where an attacker would manually hit
- 14:05the enter key for each individual
- 14:07communication, there is certainly a
- 14:09difference in volume. However, even in
- 14:13the past, attackers created their own
- 14:16attack programs to run in a loop 24/7,
- 14:19so when compared to that, it’s
- 14:21becoming difficult to tell whether the
- 14:24source is actually AI or just a script.
- 14:26Oh, I see, so it's impossible to tell.
- 14:30Wow.
- 14:31>> Also, one last thing I’d like to ask
- 14:33is about the industries I’ve been
- 14:35seeing in recent data breaches. Yes.
- 14:38Yes.
- 14:39>> In the past, it felt like the targets
- 14:41were organizations with high potential
- 14:43returns, such as large-scale businesses
- 14:46with tens of millions of customer
- 14:47records containing personal information
- 14:49, or infrastructure-related entities.
- 14:52>> I feel like those were the cases that
- 14:54were targeted most often.
- 14:55>> But quite a lot of the recent public
- 14:58reports involve
- 14:59>> places like high schools, for example,
- 15:01or businesses that don't have an
- 15:03enormous customer base. Is this because
- 15:07the quality of attacks has improved due
- 15:10to AI, for instance?
- 15:12>> I get the impression that they are now
- 15:14targeting things that humans wouldn't
- 15:16have bothered with before; what's going
- 15:18on there?
- 15:18>> Well, your impression is basically
- 15:21correct. With AI, you can target a
- 15:23wider range of victims. While some
- 15:27targets may share common IT
- 15:29infrastructure, every company and
- 15:31organization has different systems, so
- 15:33it used to be that a "one-size-fits-all
- 15:36" attack couldn't easily take everyone
- 15:38down. On the other hand, with AI, you
- 15:42can send a test attack request, look at
- 15:44the response, and realize, "If I change
- 15:47this part based on that response, it
- 15:49might get through." Previously, humans
- 15:51had to rack their brains and spend time
- 15:54on this, but now AI automatically
- 15:56decides, "Let's make the next request
- 15:58like this." I think the current
- 16:00situation is that they can now launch a
- 16:03wide range of customized attacks on a
- 16:05much broader scale.
- 16:06>> Yeah. Companies that were never
- 16:08targeted before, even those with only a
- 16:10few dozen customers, could increasingly
- 16:12become targets.
- 16:13>> That's right. To avoid any
- 16:15misunderstanding, there have always
- 16:17been broad-based attack tactics, but
- 16:20because there wasn't time to customize
- 16:22them, the probability of them hitting
- 16:24the mark was low. On the other hand,
- 16:27for large companies, customizing
- 16:28attacks could be worth it if successful
- 16:30, so they have been creating attacks
- 16:32that are somewhat more effective.
- 16:34That’s why there were so many leaks
- 16:36from major companies. That was the case
- 16:38, but since AI can now automate
- 16:39customization for smaller targets as
- 16:41well, I think it’s become easier for
- 16:43a wider range of leak incidents to
- 16:45occur, which is the structure we're
- 16:47seeing.
- 16:48>> Yeah.
- 16:49>> I see.
- 16:50>> It hasn’t been indicated that this AI
- 16:53was used in the recent domestic
- 16:55incident.
- 16:56>> However, if they were to use AI,
- 16:59>> what specific things could attackers do
- 17:01?
- 17:02>> Well, let’s see. I’d like to move
- 17:05to the next slide, but, well, how
- 17:07should I put it? It’s not really that
- 17:09a bunch of attacks through completely
- 17:12different paths than before are being
- 17:14created just because AI exists. At this
- 17:16stage, attacks that find and exploit
- 17:19vulnerabilities in public websites and
- 17:21APIs have existed since the 2010s. And,
- 17:24well, attacks targeting things like
- 17:26VPNs have also existed since the 2010s.
- 17:28Well, regarding the entry points for
- 17:31these kinds of attacks, where humans
- 17:33used to exploit vulnerabilities or send
- 17:36phishing emails to steal passwords, the
- 17:39current situation is that AI has
- 17:41largely automated and codified those
- 17:44processes.
- 17:44>> Mm-hmm.
- 17:46>> So, it's not that entirely new methods
- 17:48have been created, but rather that it
- 17:49has become easier to break through.
- 17:51>> That’s right. For example, if we look
- 17:54at websites and APIs, so-called
- 17:56unauthorized access often stems from
- 17:59vulnerabilities inherent in those
- 18:01websites or APIs. These might be unique
- 18:04vulnerabilities created by each company
- 18:06, or they could be vulnerabilities in
- 18:09common components known as libraries
- 18:11used by many. In any case, attackers
- 18:14would find these, probe them, and carry
- 18:17out attacks to achieve their objectives
- 18:20. With AI, the tasks of finding those
- 18:24vulnerabilities and then building the
- 18:26attack programs based on them......or
- 18:29creating the program and fine-tuning it
- 18:31until it actually works, all of this
- 18:33has now become automatable. As I
- 18:36mentioned regarding the MITHOS case,
- 18:38while there are many people who can
- 18:40find vulnerabilities, the number of
- 18:42people who can actually build an attack
- 18:44to exploit them is quite limited.
- 18:46It’s technically difficult. However,
- 18:48since AI is becoming quite capable of
- 18:51doing that, it has become much easier
- 18:53to launch attacks on public websites,
- 18:56which is one major situation we are
- 18:58facing.
- 18:59>> Ah, I suppose we also have to consider
- 19:02cyberattacks originating from employee
- 19:04accounts or devices.
- 19:06>> Exactly. That’s right; while
- 19:09vulnerability-based attacks have always
- 19:11been scary, the most common attack path
- 19:13for many years has actually been the
- 19:16theft of credentials, followed by
- 19:18impersonation to gain unauthorized
- 19:20access. So, there are various ways to
- 19:22steal authentication credentials, but
- 19:24the typical path is phishing. For
- 19:27example, via email or messaging apps,
- 19:29they might impersonate Apple, telling
- 19:31you your password is about to expire
- 19:33and asking you to update it. You’re
- 19:35then led to an Apple-looking site where
- 19:37you enter your ID and password, only to
- 19:40realize it was a fake site and you’ve
- 19:42just handed your Apple password over to
- 19:44a malicious party. These kinds of
- 19:46incidents have been around for a long
- 19:48time.
- 19:49>> But now, they can craft messages that
- 19:51sound more human to lure you in, or
- 19:53create phishing sites that look
- 19:55incredibly authentic. They can increase
- 19:57the volume of these attacks and improve
- 20:00their quality as well.
- 20:01>> That’s one aspect that has emerged.
- 20:03>> This next part might be a bit intense,
- 20:05so it might get cut from the broadcast,
- 20:07but I’ll go ahead and share it. There
- 20:09have also been attacks where your
- 20:11device gets taken over just by clicking
- 20:13a link. They’ve existed for a while,
- 20:15but they weren't very common. However,
- 20:17AI has made it much easier to create
- 20:20these types of attacks that take
- 20:22control just from a single link click.
- 20:24>> So, it’s becoming much more common to
- 20:26have your device hijacked simply
- 20:27because you clicked a link in a
- 20:29phishing message.
- 20:29>> Wow.
- 20:30>> Oh, is that right?
- 20:32>> Just with that one touch?
- 20:33>> Exactly, exactly, exactly. Is this like
- 20:36a computer exploit?
- 20:37>> Well, it’s a bit different from a
- 20:39standard exploit, but when you click a
- 20:41link, it has to pass through various
- 20:43pieces of software to get from your
- 20:45email app to the destination. For
- 20:46mobile apps, there are things called "
- 20:49deep links," like the ones that open a
- 20:51specific app like LINE, right? When you
- 20:53click that, the control goes to the OS,
- 20:56like Android, and if they exploit a
- 20:58vulnerability in any of the software
- 21:00involved—like the app that eventually
- 21:03opens or the browser—they can hijack
- 21:05some level of control, depending on
- 21:07which vulnerability is exploited. This
- 21:10has actually been happening for a long
- 21:12time.
- 21:13>> That’s the situation. From your
- 21:16perspective, Naka-san, are these
- 21:18strings of attacks being carried out by
- 21:21......
- 21:21>> some famous hacker group that has
- 21:23acquired incredibly innovative
- 21:25technology?
- 21:25>> Yeah. Yeah. Yeah.
- 21:28>> So, are they targeting specific areas,
- 21:30or is the number of attackers
- 21:31themselves gradually increasing?
- 21:32>> Yes. How do you see the situation, with
- 21:35various people launching various types
- 21:36of attacks?
- 21:37>> Well, I do expect that the number of
- 21:39attackers and organizations is likely
- 21:41increasing, and I’d like to discuss
- 21:44this from two perspectives. The first
- 21:46is how much easier attack tools have
- 21:49become for attackers to use. We’ve
- 21:51talked about Frontier AI and GLM-5.3,
- 21:54and it’s become easier to acquire
- 21:57high-performance AI itself. Furthermore
- 22:01, open-source software specifically
- 22:03designed for cyberattacks is
- 22:05circulating as well. While companies
- 22:10like ours monitor such tools from a
- 22:12technical standpoint, they can
- 22:14unfortunately be used for malicious
- 22:16purposes, so we are in a situation
- 22:19where both AI for easy attacks and the
- 22:21surrounding harness tools have become
- 22:23easily accessible. That’s the first
- 22:27point.
- 22:27>> Yeah. Yeah.
- 22:29>> Also, looking at the damage status—
- 22:31even just looking at the situation in
- 22:34Japan—companies are being hit for
- 22:36various reasons. For example, at one
- 22:40company, their cloud service account
- 22:42was compromised through phishing, and
- 22:44files in their drive app were accessed.
- 22:51There are leaks like that, and there
- 22:53are also cases where multiple companies
- 22:55using the same outsourced inquiry form
- 22:58service had that service provider
- 23:00attacked, leading to the leakage of
- 23:02customer names and email addresses
- 23:04entered via those forms. We’re also
- 23:10seeing instances where companies' own
- 23:13applications are being breached. Since
- 23:17they are being hit by various factors,
- 23:20I don't believe it's just one
- 23:22organization using the same AI to carry
- 23:24out these attacks.
- 23:26>> Yeah.
- 23:27>> Now, we tend to focus our attention on
- 23:30domestic leases or the trends of
- 23:32domestic companies, but...
- 23:34>> How is it globally? Is it also
- 23:36increasing worldwide since this autumn?
- 23:38>> Hmm. That’s right. It seems to be
- 23:41increasing globally. While there isn't
- 23:44much hard evidence or data yet, one
- 23:47reference point is Check Point, an
- 23:49Israeli company that works in security
- 23:52—specifically network and cloud
- 23:54security. They are in a position to
- 24:00monitor what kind of traffic is coming
- 24:02into their customers who use Check
- 24:04Point services, and whether
- 24:06cyberattacks are on the rise.
- 24:11>> According to their report, an increase
- 24:14in cyberattacks has been observed since
- 24:16September, with the number of attacks
- 24:19across all industries in September up
- 24:2148%compared to last September.
- 24:25>> Yeah. Narrowing it down to the
- 24:28financial sector, they reported a 14%
- 24:31month-on-month increase from August to
- 24:33September, so the data suggests that
- 24:36attacks are increasing not just in
- 24:38Japan, but globally. Yeah. And
- 24:42domestically, specifically regarding
- 24:44the TimeZukan case, the leak of
- 24:46numerous driver's license images has
- 24:48become a major issue, hasn't it? Do you
- 24:50think there’s a risk that AI could be
- 24:52misused here as well? Ah, yes.
- 24:55Everything we've discussed so far has
- 24:58been about how AI is used to break in.
- 25:00But if you ask what can be done with
- 25:02the information obtained using AI, the
- 25:04easiest thing to imagine is deepfakes.
- 25:06>> Exactly.
- 25:07>> Mr. Nakatani, you’ve actually been
- 25:09involved in the development of eKYC,
- 25:12but "eKYC" isn't a very familiar term,
- 25:14is it?
- 25:15>> Yes. That's right. Is that identity
- 25:17verification?
- 25:18>> Exactly. You know how, lately, on many
- 25:21sites when you register as a user,
- 25:23>> Yes.
- 25:24>> they ask you to show your driver's
- 25:25license,
- 25:26>> or to look this way. Look to the side.
- 25:28>> They do that, right? Or they ask you to
- 25:29show the thickness of your license. Ah,
- 25:31that’s the one. So, KYC. Yeah.
- 25:34>> I wondered what KYC meant, so I looked
- 25:35it up, and it’s "Know Your Customer."
- 25:37>> Yes. Ah.
- 25:38>> It's actually quite simple.
- 25:40>> Like, about you.
- 25:41>> Like, knowing your customer. Well,
- 25:43it’s electronic, or, wait, e,
- 25:45electronic, that's where the 'e' comes
- 25:48from. So, that's the 'c', but is there
- 25:51a possibility that it could be easily
- 25:54bypassed using AI?
- 25:55>> That's right. I actually think so. The
- 25:58current eKYC system in Japan is
- 26:01basically based on the Act on
- 26:03Prevention of Transfer of Criminal
- 26:05Proceeds, the so-called "Anti-Criminal
- 26:08Proceeds Act," and it uses methods like
- 26:10the "i-ro-ha" approach. As you
- 26:13mentioned earlier, this relies on
- 26:15looking at the face of the ID document,
- 26:18checking the appearance and the
- 26:20thickness to verify it's not a forged
- 26:22document, or at least that's the logic.
- 26:25And on top of that, they check the
- 26:27person's face, not just a still image,
- 26:30but by having them show movement to
- 26:32confirm they are real. That’s the
- 26:34method they use. However, I think this
- 26:37technology and the system itself were
- 26:40designed before AI existed. As you said
- 26:43, I also suspect it could be easily
- 26:46bypassed with deepfakes. My reasoning
- 26:49is that, as you mentioned earlier, I
- 26:51used to work at a company that
- 26:53developed consumer-facing services.
- 26:56While we didn't build the eKYC systems
- 26:59ourselves, I was the project leader
- 27:01responsible for implementing eKYC to
- 27:04verify the identities of customers
- 27:06using our company’s application.
- 27:08>> Uh-huh. Yeah.
- 27:11>> Because of that, I approached several
- 27:13vendors who build these systems, and
- 27:16while we were in the process of
- 27:17selecting one, one company...
- 27:21>> Regarding that eKYC system, we wanted
- 27:24to test it from our perspective. When
- 27:29testing, using our own developer IDs or
- 27:32facial photos raised privacy concerns
- 27:35for our employees. So, we decided to
- 27:40buy party masks—like a Donald Trump
- 27:43mask or one of then-Prime Minister Abe
- 27:46—and create driver's license images
- 27:48wearing those masks. Of course, these
- 27:52were documents made solely for testing
- 27:55purposes. We printed them out, verified
- 27:57they passed the thickness check, and
- 27:59when we followed the instructions to
- 28:01turn right while wearing the mask, it
- 28:03actually cleared the verification
- 28:05process.
- 28:06>> Wow. Granted, this was back in 2018 or
- 28:102019, and while companies have likely
- 28:13improved their technical accuracy
- 28:15within legal frameworks since then,
- 28:17generative AI and deepfakes can now
- 28:19create much more precise facial images
- 28:22than those masks. I believe today's
- 28:25methods have their limits. Oh.
- 28:27>> Ah,
- 28:28>> Yeah.
- 28:29>> And at a time like this, Mr. Takezaki
- 28:31...
- 28:31>> Yes. Our time has already flown by. I'm
- 28:34sorry. I learned so much. No, thank you
- 28:37as well.
- 28:38>> Please, keep asking him plenty of
- 28:39questions.
- 28:40>> I have to head to another job, sorry.
- 28:42Excuse me.
- 28:43>> Good work today.
- 28:46>> I'll just move this chair.
- 28:48>> I recall these discussions heating up
- 28:51in the U.S., and I feel like the OpenAI
- 28:53and Hugging Face situation was really
- 28:56the trigger for it.
- 28:57>> And that, well, was an example of AI
- 29:00running wild, wasn't it? Should we view
- 29:03this current incident as strictly the
- 29:05work of malicious individuals?
- 29:08>> Or can we assume there is zero
- 29:10possibility that an AI went rogue and
- 29:12orchestrated this? Or should we
- 29:15consider that there might be a slight
- 29:17chance of that? What do you think?
- 29:19>> This is an extremely good and yet
- 29:22extremely difficult question, so I will
- 29:25answer it carefully. I think that as
- 29:28long as we, the Japanese and Americans,
- 29:31are using well-known and widely used AI
- 29:34like OpenAI or Anthropic, it is
- 29:36unlikely that this kind of malicious
- 29:38runaway behavior will happen. This is
- 29:42because, as I mentioned earlier, the
- 29:44guardrails are...
About this transcript
This page contains the full transcript of 【情報漏洩】中国製AI「GLM」の悪用か?タイムズなど大手の被害続発・・・「最新AI」の脅威と自分の情報を守る「3つの防衛策」 by テレ東BIZ ダイジェスト, generated from the public captions YouTube serves with the video. The transcript has 4,557 words across 780 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.