YouTube2Text

【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

by テレ東BIZ ダイジェスト · 4,557 words · 780 segments · language auto · Watch on YouTube

Full transcript

  1. 0:00Hello, everyone. Today, we will take a
  2. 0:03deep dive into the cyberattacks
  3. 0:05targeting companies and the massive
  4. 0:08data breaches being reported daily.
  5. 0:10Starting with the data breach of up to
  6. 0:12approximately 6.6 million records at
  7. 0:14Times Car, announcements of damages
  8. 0:15targeting companies and organizations
  9. 0:17have been appearing one after another
  10. 0:19over the past few days. Joining us in
  11. 0:21the studio today are Mr. Sho Nakatani
  12. 0:23from AgenticSec, an expert in AI within
  13. 0:26the field of cybersecurity, and Ms. Umi
  14. 0:28of TV Tokyo, who covers the technology
  15. 0:30sector. It’s a pleasure to have you
  16. 0:33both with us.
  17. 0:33>> Thank you for having us.
  18. 0:35>> You appeared in our previous
  19. 0:36explanatory video for Claude Multus,
  20. 0:38didn't you? Yes. Mr. Nakatani, you run
  21. 0:41a security business, don't you?
  22. 0:43>> Yes. That's right. At AgenticSec, we
  23. 0:46provide products that help protect our
  24. 0:49customers 'cybersecurity by using the
  25. 0:52same techniques attackers use to stay
  26. 0:54one step ahead and identify risks early
  27. 0:57.
  28. 0:57>> Uh-huh. I see, so you are providing
  29. 1:00support to those companies.
  30. 1:02>> You're well-versed in the intersection
  31. 1:04of AI and security, I presume.
  32. 1:05>> That's right. Exactly. Since attackers
  33. 1:08nowadays are using AI, we offer
  34. 1:10products that simulate what happens
  35. 1:12when AI is used to carry out security
  36. 1:14attacks, so we can demonstrate those
  37. 1:17risks.
  38. 1:17>> I see. Today, I'd like you to explain
  39. 1:20how AI is being used behind the scenes.
  40. 1:23>> Certainly.
  41. 1:24>> Now, regarding the data breaches being
  42. 1:27reported day after day, their scale and
  43. 1:29scope continue to expand. At discount
  44. 1:31store operator MRIS Tarmarks, up to
  45. 1:34approximately 1.73 million records. At
  46. 1:37GMO's survey site, up to approximately
  47. 1:39950,000 records, and at Yakiniku King,
  48. 1:4210.78 million records, which accounts
  49. 1:45for nearly all of their members. It has
  50. 1:47been announced that 550,000 records at
  51. 1:50an Asahi subsidiary and 100,000 records
  52. 1:53at Citizen Watch may have been leaked.
  53. 1:55Furthermore, ransomware attacks on
  54. 1:57institutions like Daiwa Securities and
  55. 2:00Osaka University have caused immense
  56. 2:02damage across all industries. In
  57. 2:04response to this situation, Digital
  58. 2:06Minister Furukawa called on the public
  59. 2:07at a post-Cabinet meeting press
  60. 2:09conference to be conscious of
  61. 2:10protecting their own information, and
  62. 2:12the Chief Cabinet Secretary has also
  63. 2:14expressed that the government is taking
  64. 2:16this very seriously and heightening its
  65. 2:18vigilance. It seems like these
  66. 2:19incidents have surged over the past few
  67. 2:21weeks, but what is your take on this,
  68. 2:23Mr. Nakatani?
  69. 2:24>> Well, that’s true. I’m on X
  70. 2:26whenever I have a spare moment, and
  71. 2:28starting around the end of September,
  72. 2:31with the Times incident leading the way
  73. 2:33, we’ve seen news of data leaks
  74. 2:35almost every single day. As someone who
  75. 2:38runs a company dedicated to security,
  76. 2:40it truly pains me to see this happening
  77. 2:43. Yeah.
  78. 2:44>> Should we view this as a genuine
  79. 2:46increase over the last few weeks, or is
  80. 2:48it more a case of existing issues
  81. 2:49finally coming to light?
  82. 2:51>> I’d say it’s both. I believe the
  83. 2:53pattern of these issues coming to light
  84. 2:56is definitely a factor. We are seeing a
  85. 2:59significant rise in reported cases, but
  86. 3:02the timing of the actual attacks is a
  87. 3:04separate matter. Seeing so many other
  88. 3:08companies being targeted, some firms
  89. 3:10are conducting internal reviews,
  90. 3:12telling their teams to check again if
  91. 3:15they’ve been hit. I suspect that in
  92. 3:19some cases, they’ve discovered they
  93. 3:21were actually compromised months ago.
  94. 3:25So, while it is a fact that attacks are
  95. 3:27increasing, the rise in incidents we
  96. 3:30see now involves both fresh attacks and
  97. 3:32older ones that are only just now being
  98. 3:35uncovered.
  99. 3:36>> Yeah. It feels as though corporate
  100. 3:39vigilance is also intensifying.
  101. 3:42According to a report by the National
  102. 3:44Police Agency, the number of suspicious
  103. 3:47access attempts classified as the
  104. 3:49preparatory stage of a cyberattack
  105. 3:51reached 13,687 per day per IP address
  106. 3:54in the first half of this year. This is
  107. 3:57a 50%increase compared to the same
  108. 3:59period last year, marking an all-time
  109. 4:02high. Mr. Nakatani, personally, I
  110. 4:04suspect that the latest AI technology
  111. 4:07is driving this increase in
  112. 4:09cyberattacks. What is your actual view
  113. 4:12on that?
  114. 4:14>> I personally share the same view; as AI
  115. 4:16models have continued to advance, I
  116. 4:19definitely feel there is a correlation
  117. 4:21with the current rise in attacks.
  118. 4:24Mm-hmm.
  119. 4:25>> However, it’s not yet clear whether
  120. 4:28the increase in attacks we're seeing
  121. 4:30now is definitively due to these newer
  122. 4:32models. There isn't enough evidence
  123. 4:34circulating in the world to make a
  124. 4:36definitive judgment just yet.
  125. 4:38>> Mm-hmm.
  126. 4:41>> Well, speaking of whether or not
  127. 4:43state-of-the-art AI models are being
  128. 4:45used, what I’m focusing on right now
  129. 4:47are open-weight models. These include
  130. 4:51open-weight models often produced in
  131. 4:53places like China, and one that's been
  132. 4:56a hot topic in the security community
  133. 4:58lately is the GLM 5.3 model. Compared
  134. 5:03to something like Anthropic's Mythos
  135. 5:06released in April, this model has a
  136. 5:08comparable level of cybersecurity
  137. 5:10attack capability, yet it lacks the
  138. 5:13guardrails that US frontier models—
  139. 5:16like Mythos, Fable, or Astra GPT-6 Sol
  140. 5:18—have to prevent dangerous
  141. 5:20cyberattacks. Detailed reports from
  142. 5:26Anthropic suggest that the rise in
  143. 5:29cyberattacks may be linked to the
  144. 5:31emergence of models like GLM 5.3 that
  145. 5:34possess high capabilities without those
  146. 5:36built-in guardrails. I personally
  147. 5:39believe this is likely a contributing
  148. 5:41factor.
  149. 5:42>> Mm-hmm.
  150. 5:43>> Is GLM 5.3 really that powerful?
  151. 5:46>> It is incredibly impressive, yes. We
  152. 5:48saw reports about how cybersecurity
  153. 5:50capabilities had shifted and improved
  154. 5:52drastically with Mythos back in April,
  155. 5:54and here we are in October, so in just
  156. 5:57about six months, that level of power
  157. 5:59has become available in an open-weight
  158. 6:01format. The fact that something
  159. 6:03appeared that is easy for attackers to
  160. 6:06use because it lacks guardrails was
  161. 6:08something we anticipated, but it
  162. 6:10happened faster than expected. Combined
  163. 6:13with the rise in these attacks, I feel
  164. 6:16that while the impact is as large as we
  165. 6:18imagined, it is also proving to be even
  166. 6:21more concerning than I had anticipated.
  167. 6:24>> Mm-hmm. To add a bit of context,
  168. 6:27regarding the point about "open weights
  169. 6:29," you have models like the well-known
  170. 6:32ChatGPT or the Claude series from
  171. 6:34Anthropic—which we've mentioned—
  172. 6:36that are considered closed models, and
  173. 6:39then there is the open-source category,
  174. 6:41but on the other hand, "
  175. 6:43>> Open weights" sits somewhere in between
  176. 6:45. In other words, since the weights are
  177. 6:48something users can tinker with a bit
  178. 6:50—whereas changing the settings of
  179. 6:52ChatGPT itself is difficult—with open
  180. 6:55weights, users can tune the model to
  181. 6:57their own preferences, is that right?
  182. 7:00>> Yes, that’s correct. Well, with an
  183. 7:02open-weight model, you can obviously
  184. 7:04just download it to your own
  185. 7:06environment and use it on a high-spec
  186. 7:08machine without tuning it. Plus, by
  187. 7:10toggling the weights—the connection
  188. 7:12values between the model’s parameters
  189. 7:14that you mentioned—you can, for
  190. 7:15instance, adjust the alignment, which
  191. 7:17is the part that tells the model not to
  192. 7:19do anything dangerous from a security
  193. 7:21standpoint. So, it is technically
  194. 7:23possible to tamper with those sorts of
  195. 7:26guardrails. For GLM-4, it currently
  196. 7:30doesn’t seem like there’s much need
  197. 7:32to go through the trouble of modifying
  198. 7:34those weights, but in the future, if
  199. 7:36guardrails and alignment for
  200. 7:38open-weight models become too strict,
  201. 7:40it wouldn't be surprising to see people
  202. 7:43take advantage of the open weights to
  203. 7:45modify them and potentially misuse the
  204. 7:47model. So, the extent of the guardrails
  205. 7:51provided is essentially left up to the
  206. 7:54morality of the development company,
  207. 7:56isn't it? I believe it really comes
  208. 7:58down to their morals, what they are
  209. 8:00aiming to achieve with their
  210. 8:02development, and what kind of users
  211. 8:04they want to reach.
  212. 8:05>> Well, for example, regarding Zhipu AI,
  213. 8:08the Chinese company behind GLM-4, I
  214. 8:10don't know the exact intentions behind
  215. 8:13their release. However, compared to US
  216. 8:16companies like OpenAI or Anthropic,
  217. 8:19performance remains their top priority
  218. 8:21—their primary focus is on catching
  219. 8:24up to and surpassing the US. Meanwhile,
  220. 8:27US companies have already reached a
  221. 8:30point where, in terms of cybersecurity
  222. 8:32and mathematical capabilities—at
  223. 8:34least within the scope of what
  224. 8:36benchmarks can measure—they've gone
  225. 8:38quite far. Instead of pushing those
  226. 8:42further and competing over
  227. 8:43harder-to-understand benchmarks, I
  228. 8:46believe US firms are now shifting their
  229. 8:48focus toward safety and broadening
  230. 8:51their general capabilities beyond just
  231. 8:53cyber and math.
  232. 8:54>> Ah, I see. So it’s your impression
  233. 8:58that China is still heavily focused on
  234. 9:00growth in the security sector?
  235. 9:01>> I can't say for sure since I don't have
  236. 9:04exact information, but regarding the
  237. 9:06GLM-4 model, I definitely feel there's
  238. 9:08a level of performance growth that at
  239. 9:10least suggests security capabilities
  240. 9:13are being targeted.
  241. 9:14>> Looking at it from your perspective,
  242. 9:15Mr. Nakatani, it's only been a few
  243. 9:17months since the release of the
  244. 9:18previous model, right?
  245. 9:19>> Yes. Yes. So, was it somewhat expected
  246. 9:25that China would release a model like
  247. 9:28this GLM that rivals frontier models,
  248. 9:30especially one with open weights?
  249. 9:35>> That's correct. I did expect it. As I
  250. 9:38mentioned in a previous interview,
  251. 9:40there has consistently been a trend
  252. 9:43where open-weight models catch up in
  253. 9:45performance a few months behind US
  254. 9:47frontier models like those from OpenAI
  255. 9:50or Anthropic—it’s generally said to
  256. 9:52be a lag of about six to ten months. So
  257. 9:55, with the release of models like
  258. 9:57Claude 3, I think I mentioned somewhere
  259. 10:00that this wouldn't just be something
  260. 10:02Anthropic could do, that OpenAI would
  261. 10:04follow suit, and that open-weight
  262. 10:06models would catch up in about half a
  263. 10:09year or so. Honestly, I feel that's
  264. 10:11exactly what ended up happening.
  265. 10:13>> Yeah.
  266. 10:14>> I see.
  267. 10:15>> But what is the reason why we can't
  268. 10:17definitively conclude that this whole
  269. 10:20series of incidents was caused by
  270. 10:22things like GLM or other AI models?
  271. 10:24>> Well, let's see. I'll explain this
  272. 10:27while showing you a diagram.
  273. 10:31>> The key point in this diagram is that
  274. 10:33the victim system on the far right
  275. 10:36receives attack traffic from two
  276. 10:38sources and cannot distinguish between
  277. 10:40them. These two sources, as shown,
  278. 10:44represent an AI-driven cyberattack at
  279. 10:47the top and a conventional, non-AI
  280. 10:50cyberattack at the bottom. Saying that
  281. 10:52they are indistinguishable...is what
  282. 10:54this illustration shows.
  283. 10:56>> Yes.
  284. 10:57>> Furthermore, if AI attacks and
  285. 10:59conventional attacks are
  286. 11:01indistinguishable, it's even more
  287. 11:03impossible to tell which AI was used
  288. 11:05for an AI-based attack. Therefore, I'm
  289. 11:08saying it's honestly difficult for the
  290. 11:10victim side to identify which model was
  291. 11:13used. What exactly is a harness?
  292. 11:16>> A harness is...well, before talking
  293. 11:18about the harness, if you look at the
  294. 11:20AI model on the left, this is the basic
  295. 11:23behavior you experience when talking to
  296. 11:25something like ChatGPT—you ask a
  297. 11:27question in a language, and a language
  298. 11:30response comes back. This is
  299. 11:32essentially the AI model receiving
  300. 11:34tokens and outputting tokens. But if it
  301. 11:39only outputs language, it's hard to
  302. 11:42perform things like cyberattacks, or in
  303. 11:45some cases, generating diagrams or
  304. 11:48images, so then, what exactly is...
  305. 11:52>> When giving a model the ability to
  306. 11:55execute computer programs, we attach
  307. 11:57what's called a harness to it;
  308. 11:59basically, it's a tool or a layer that
  309. 12:01enables the execution of code or source
  310. 12:04code.
  311. 12:04>> Mm. Yeah.
  312. 12:07>> So, there’s currently a two-step
  313. 12:09process: the AI model generates
  314. 12:11instructions, and then this control
  315. 12:13infrastructure, or harness, executes
  316. 12:15them.
  317. 12:18>> And you don't even have to consciously
  318. 12:20think about passing the model's output
  319. 12:22to the harness anymore; if you use
  320. 12:24tools like the famous coding agents
  321. 12:27Codex or Claude Code, the user just
  322. 12:29sends an instruction, it automatically
  323. 12:31reaches the model, the model processes
  324. 12:34the tokens, and then it calls the
  325. 12:39harness as needed to perform the action
  326. 12:42the user wants—that's how AI works
  327. 12:44now.
  328. 12:45>> Mm.
  329. 12:46>> If guardrails are in place, does that
  330. 12:48mean the AI model simply won't output
  331. 12:50the code to do those things in the
  332. 12:52first place?
  333. 12:53>> It’s actually both; there are parts
  334. 12:55prevented on the model side and parts
  335. 12:57prevented on the harness side. The
  336. 13:00prevention on the model side uses a
  337. 13:03technique often called alignment; when
  338. 13:05the model spits out tokens, it has been
  339. 13:08fine-tuned—or "aligned"—so that
  340. 13:10it's embedded with instructions like "
  341. 13:12don't say things like this," making it
  342. 13:15so those tokens generally don't even
  343. 13:17come out. But since that alone is a bit
  344. 13:20too weak as a guarantee, you can also
  345. 13:22have the harness block things like a
  346. 13:24blacklist, ensuring certain actions are
  347. 13:26never taken, and that collection of
  348. 13:28controls is what we call "rails." Mm.
  349. 13:31Yeah.
  350. 13:32>> Now, as for this GLM, while it does
  351. 13:35have guardrails, they are, in a sense,
  352. 13:37very low.
  353. 13:38>> I think that’s the situation.
  354. 13:40>> I see. Mm.
  355. 13:42>> I’d like to ask one more thing:
  356. 13:43regarding attacks from the AI side, or
  357. 13:46attacks using tools, or attacks by
  358. 13:47humans—even if you can't distinguish
  359. 13:49between them, for example...
  360. 13:51>> It's easy to imagine that with AI, you
  361. 13:53could just keep sending attacks
  362. 13:55indefinitely, like 24/7, so wouldn't
  363. 13:57there be a difference in volume? Is
  364. 13:59that not the case?
  365. 14:00>> Yes. Compared to old-school attacks
  366. 14:03where an attacker would manually hit
  367. 14:05the enter key for each individual
  368. 14:07communication, there is certainly a
  369. 14:09difference in volume. However, even in
  370. 14:13the past, attackers created their own
  371. 14:16attack programs to run in a loop 24/7,
  372. 14:19so when compared to that, it’s
  373. 14:21becoming difficult to tell whether the
  374. 14:24source is actually AI or just a script.
  375. 14:26Oh, I see, so it's impossible to tell.
  376. 14:30Wow.
  377. 14:31>> Also, one last thing I’d like to ask
  378. 14:33is about the industries I’ve been
  379. 14:35seeing in recent data breaches. Yes.
  380. 14:38Yes.
  381. 14:39>> In the past, it felt like the targets
  382. 14:41were organizations with high potential
  383. 14:43returns, such as large-scale businesses
  384. 14:46with tens of millions of customer
  385. 14:47records containing personal information
  386. 14:49, or infrastructure-related entities.
  387. 14:52>> I feel like those were the cases that
  388. 14:54were targeted most often.
  389. 14:55>> But quite a lot of the recent public
  390. 14:58reports involve
  391. 14:59>> places like high schools, for example,
  392. 15:01or businesses that don't have an
  393. 15:03enormous customer base. Is this because
  394. 15:07the quality of attacks has improved due
  395. 15:10to AI, for instance?
  396. 15:12>> I get the impression that they are now
  397. 15:14targeting things that humans wouldn't
  398. 15:16have bothered with before; what's going
  399. 15:18on there?
  400. 15:18>> Well, your impression is basically
  401. 15:21correct. With AI, you can target a
  402. 15:23wider range of victims. While some
  403. 15:27targets may share common IT
  404. 15:29infrastructure, every company and
  405. 15:31organization has different systems, so
  406. 15:33it used to be that a "one-size-fits-all
  407. 15:36" attack couldn't easily take everyone
  408. 15:38down. On the other hand, with AI, you
  409. 15:42can send a test attack request, look at
  410. 15:44the response, and realize, "If I change
  411. 15:47this part based on that response, it
  412. 15:49might get through." Previously, humans
  413. 15:51had to rack their brains and spend time
  414. 15:54on this, but now AI automatically
  415. 15:56decides, "Let's make the next request
  416. 15:58like this." I think the current
  417. 16:00situation is that they can now launch a
  418. 16:03wide range of customized attacks on a
  419. 16:05much broader scale.
  420. 16:06>> Yeah. Companies that were never
  421. 16:08targeted before, even those with only a
  422. 16:10few dozen customers, could increasingly
  423. 16:12become targets.
  424. 16:13>> That's right. To avoid any
  425. 16:15misunderstanding, there have always
  426. 16:17been broad-based attack tactics, but
  427. 16:20because there wasn't time to customize
  428. 16:22them, the probability of them hitting
  429. 16:24the mark was low. On the other hand,
  430. 16:27for large companies, customizing
  431. 16:28attacks could be worth it if successful
  432. 16:30, so they have been creating attacks
  433. 16:32that are somewhat more effective.
  434. 16:34That’s why there were so many leaks
  435. 16:36from major companies. That was the case
  436. 16:38, but since AI can now automate
  437. 16:39customization for smaller targets as
  438. 16:41well, I think it’s become easier for
  439. 16:43a wider range of leak incidents to
  440. 16:45occur, which is the structure we're
  441. 16:47seeing.
  442. 16:48>> Yeah.
  443. 16:49>> I see.
  444. 16:50>> It hasn’t been indicated that this AI
  445. 16:53was used in the recent domestic
  446. 16:55incident.
  447. 16:56>> However, if they were to use AI,
  448. 16:59>> what specific things could attackers do
  449. 17:01?
  450. 17:02>> Well, let’s see. I’d like to move
  451. 17:05to the next slide, but, well, how
  452. 17:07should I put it? It’s not really that
  453. 17:09a bunch of attacks through completely
  454. 17:12different paths than before are being
  455. 17:14created just because AI exists. At this
  456. 17:16stage, attacks that find and exploit
  457. 17:19vulnerabilities in public websites and
  458. 17:21APIs have existed since the 2010s. And,
  459. 17:24well, attacks targeting things like
  460. 17:26VPNs have also existed since the 2010s.
  461. 17:28Well, regarding the entry points for
  462. 17:31these kinds of attacks, where humans
  463. 17:33used to exploit vulnerabilities or send
  464. 17:36phishing emails to steal passwords, the
  465. 17:39current situation is that AI has
  466. 17:41largely automated and codified those
  467. 17:44processes.
  468. 17:44>> Mm-hmm.
  469. 17:46>> So, it's not that entirely new methods
  470. 17:48have been created, but rather that it
  471. 17:49has become easier to break through.
  472. 17:51>> That’s right. For example, if we look
  473. 17:54at websites and APIs, so-called
  474. 17:56unauthorized access often stems from
  475. 17:59vulnerabilities inherent in those
  476. 18:01websites or APIs. These might be unique
  477. 18:04vulnerabilities created by each company
  478. 18:06, or they could be vulnerabilities in
  479. 18:09common components known as libraries
  480. 18:11used by many. In any case, attackers
  481. 18:14would find these, probe them, and carry
  482. 18:17out attacks to achieve their objectives
  483. 18:20. With AI, the tasks of finding those
  484. 18:24vulnerabilities and then building the
  485. 18:26attack programs based on them......or
  486. 18:29creating the program and fine-tuning it
  487. 18:31until it actually works, all of this
  488. 18:33has now become automatable. As I
  489. 18:36mentioned regarding the MITHOS case,
  490. 18:38while there are many people who can
  491. 18:40find vulnerabilities, the number of
  492. 18:42people who can actually build an attack
  493. 18:44to exploit them is quite limited.
  494. 18:46It’s technically difficult. However,
  495. 18:48since AI is becoming quite capable of
  496. 18:51doing that, it has become much easier
  497. 18:53to launch attacks on public websites,
  498. 18:56which is one major situation we are
  499. 18:58facing.
  500. 18:59>> Ah, I suppose we also have to consider
  501. 19:02cyberattacks originating from employee
  502. 19:04accounts or devices.
  503. 19:06>> Exactly. That’s right; while
  504. 19:09vulnerability-based attacks have always
  505. 19:11been scary, the most common attack path
  506. 19:13for many years has actually been the
  507. 19:16theft of credentials, followed by
  508. 19:18impersonation to gain unauthorized
  509. 19:20access. So, there are various ways to
  510. 19:22steal authentication credentials, but
  511. 19:24the typical path is phishing. For
  512. 19:27example, via email or messaging apps,
  513. 19:29they might impersonate Apple, telling
  514. 19:31you your password is about to expire
  515. 19:33and asking you to update it. You’re
  516. 19:35then led to an Apple-looking site where
  517. 19:37you enter your ID and password, only to
  518. 19:40realize it was a fake site and you’ve
  519. 19:42just handed your Apple password over to
  520. 19:44a malicious party. These kinds of
  521. 19:46incidents have been around for a long
  522. 19:48time.
  523. 19:49>> But now, they can craft messages that
  524. 19:51sound more human to lure you in, or
  525. 19:53create phishing sites that look
  526. 19:55incredibly authentic. They can increase
  527. 19:57the volume of these attacks and improve
  528. 20:00their quality as well.
  529. 20:01>> That’s one aspect that has emerged.
  530. 20:03>> This next part might be a bit intense,
  531. 20:05so it might get cut from the broadcast,
  532. 20:07but I’ll go ahead and share it. There
  533. 20:09have also been attacks where your
  534. 20:11device gets taken over just by clicking
  535. 20:13a link. They’ve existed for a while,
  536. 20:15but they weren't very common. However,
  537. 20:17AI has made it much easier to create
  538. 20:20these types of attacks that take
  539. 20:22control just from a single link click.
  540. 20:24>> So, it’s becoming much more common to
  541. 20:26have your device hijacked simply
  542. 20:27because you clicked a link in a
  543. 20:29phishing message.
  544. 20:29>> Wow.
  545. 20:30>> Oh, is that right?
  546. 20:32>> Just with that one touch?
  547. 20:33>> Exactly, exactly, exactly. Is this like
  548. 20:36a computer exploit?
  549. 20:37>> Well, it’s a bit different from a
  550. 20:39standard exploit, but when you click a
  551. 20:41link, it has to pass through various
  552. 20:43pieces of software to get from your
  553. 20:45email app to the destination. For
  554. 20:46mobile apps, there are things called "
  555. 20:49deep links," like the ones that open a
  556. 20:51specific app like LINE, right? When you
  557. 20:53click that, the control goes to the OS,
  558. 20:56like Android, and if they exploit a
  559. 20:58vulnerability in any of the software
  560. 21:00involved—like the app that eventually
  561. 21:03opens or the browser—they can hijack
  562. 21:05some level of control, depending on
  563. 21:07which vulnerability is exploited. This
  564. 21:10has actually been happening for a long
  565. 21:12time.
  566. 21:13>> That’s the situation. From your
  567. 21:16perspective, Naka-san, are these
  568. 21:18strings of attacks being carried out by
  569. 21:21......
  570. 21:21>> some famous hacker group that has
  571. 21:23acquired incredibly innovative
  572. 21:25technology?
  573. 21:25>> Yeah. Yeah. Yeah.
  574. 21:28>> So, are they targeting specific areas,
  575. 21:30or is the number of attackers
  576. 21:31themselves gradually increasing?
  577. 21:32>> Yes. How do you see the situation, with
  578. 21:35various people launching various types
  579. 21:36of attacks?
  580. 21:37>> Well, I do expect that the number of
  581. 21:39attackers and organizations is likely
  582. 21:41increasing, and I’d like to discuss
  583. 21:44this from two perspectives. The first
  584. 21:46is how much easier attack tools have
  585. 21:49become for attackers to use. We’ve
  586. 21:51talked about Frontier AI and GLM-5.3,
  587. 21:54and it’s become easier to acquire
  588. 21:57high-performance AI itself. Furthermore
  589. 22:01, open-source software specifically
  590. 22:03designed for cyberattacks is
  591. 22:05circulating as well. While companies
  592. 22:10like ours monitor such tools from a
  593. 22:12technical standpoint, they can
  594. 22:14unfortunately be used for malicious
  595. 22:16purposes, so we are in a situation
  596. 22:19where both AI for easy attacks and the
  597. 22:21surrounding harness tools have become
  598. 22:23easily accessible. That’s the first
  599. 22:27point.
  600. 22:27>> Yeah. Yeah.
  601. 22:29>> Also, looking at the damage status—
  602. 22:31even just looking at the situation in
  603. 22:34Japan—companies are being hit for
  604. 22:36various reasons. For example, at one
  605. 22:40company, their cloud service account
  606. 22:42was compromised through phishing, and
  607. 22:44files in their drive app were accessed.
  608. 22:51There are leaks like that, and there
  609. 22:53are also cases where multiple companies
  610. 22:55using the same outsourced inquiry form
  611. 22:58service had that service provider
  612. 23:00attacked, leading to the leakage of
  613. 23:02customer names and email addresses
  614. 23:04entered via those forms. We’re also
  615. 23:10seeing instances where companies' own
  616. 23:13applications are being breached. Since
  617. 23:17they are being hit by various factors,
  618. 23:20I don't believe it's just one
  619. 23:22organization using the same AI to carry
  620. 23:24out these attacks.
  621. 23:26>> Yeah.
  622. 23:27>> Now, we tend to focus our attention on
  623. 23:30domestic leases or the trends of
  624. 23:32domestic companies, but...
  625. 23:34>> How is it globally? Is it also
  626. 23:36increasing worldwide since this autumn?
  627. 23:38>> Hmm. That’s right. It seems to be
  628. 23:41increasing globally. While there isn't
  629. 23:44much hard evidence or data yet, one
  630. 23:47reference point is Check Point, an
  631. 23:49Israeli company that works in security
  632. 23:52—specifically network and cloud
  633. 23:54security. They are in a position to
  634. 24:00monitor what kind of traffic is coming
  635. 24:02into their customers who use Check
  636. 24:04Point services, and whether
  637. 24:06cyberattacks are on the rise.
  638. 24:11>> According to their report, an increase
  639. 24:14in cyberattacks has been observed since
  640. 24:16September, with the number of attacks
  641. 24:19across all industries in September up
  642. 24:2148%compared to last September.
  643. 24:25>> Yeah. Narrowing it down to the
  644. 24:28financial sector, they reported a 14%
  645. 24:31month-on-month increase from August to
  646. 24:33September, so the data suggests that
  647. 24:36attacks are increasing not just in
  648. 24:38Japan, but globally. Yeah. And
  649. 24:42domestically, specifically regarding
  650. 24:44the TimeZukan case, the leak of
  651. 24:46numerous driver's license images has
  652. 24:48become a major issue, hasn't it? Do you
  653. 24:50think there’s a risk that AI could be
  654. 24:52misused here as well? Ah, yes.
  655. 24:55Everything we've discussed so far has
  656. 24:58been about how AI is used to break in.
  657. 25:00But if you ask what can be done with
  658. 25:02the information obtained using AI, the
  659. 25:04easiest thing to imagine is deepfakes.
  660. 25:06>> Exactly.
  661. 25:07>> Mr. Nakatani, you’ve actually been
  662. 25:09involved in the development of eKYC,
  663. 25:12but "eKYC" isn't a very familiar term,
  664. 25:14is it?
  665. 25:15>> Yes. That's right. Is that identity
  666. 25:17verification?
  667. 25:18>> Exactly. You know how, lately, on many
  668. 25:21sites when you register as a user,
  669. 25:23>> Yes.
  670. 25:24>> they ask you to show your driver's
  671. 25:25license,
  672. 25:26>> or to look this way. Look to the side.
  673. 25:28>> They do that, right? Or they ask you to
  674. 25:29show the thickness of your license. Ah,
  675. 25:31that’s the one. So, KYC. Yeah.
  676. 25:34>> I wondered what KYC meant, so I looked
  677. 25:35it up, and it’s "Know Your Customer."
  678. 25:37>> Yes. Ah.
  679. 25:38>> It's actually quite simple.
  680. 25:40>> Like, about you.
  681. 25:41>> Like, knowing your customer. Well,
  682. 25:43it’s electronic, or, wait, e,
  683. 25:45electronic, that's where the 'e' comes
  684. 25:48from. So, that's the 'c', but is there
  685. 25:51a possibility that it could be easily
  686. 25:54bypassed using AI?
  687. 25:55>> That's right. I actually think so. The
  688. 25:58current eKYC system in Japan is
  689. 26:01basically based on the Act on
  690. 26:03Prevention of Transfer of Criminal
  691. 26:05Proceeds, the so-called "Anti-Criminal
  692. 26:08Proceeds Act," and it uses methods like
  693. 26:10the "i-ro-ha" approach. As you
  694. 26:13mentioned earlier, this relies on
  695. 26:15looking at the face of the ID document,
  696. 26:18checking the appearance and the
  697. 26:20thickness to verify it's not a forged
  698. 26:22document, or at least that's the logic.
  699. 26:25And on top of that, they check the
  700. 26:27person's face, not just a still image,
  701. 26:30but by having them show movement to
  702. 26:32confirm they are real. That’s the
  703. 26:34method they use. However, I think this
  704. 26:37technology and the system itself were
  705. 26:40designed before AI existed. As you said
  706. 26:43, I also suspect it could be easily
  707. 26:46bypassed with deepfakes. My reasoning
  708. 26:49is that, as you mentioned earlier, I
  709. 26:51used to work at a company that
  710. 26:53developed consumer-facing services.
  711. 26:56While we didn't build the eKYC systems
  712. 26:59ourselves, I was the project leader
  713. 27:01responsible for implementing eKYC to
  714. 27:04verify the identities of customers
  715. 27:06using our company’s application.
  716. 27:08>> Uh-huh. Yeah.
  717. 27:11>> Because of that, I approached several
  718. 27:13vendors who build these systems, and
  719. 27:16while we were in the process of
  720. 27:17selecting one, one company...
  721. 27:21>> Regarding that eKYC system, we wanted
  722. 27:24to test it from our perspective. When
  723. 27:29testing, using our own developer IDs or
  724. 27:32facial photos raised privacy concerns
  725. 27:35for our employees. So, we decided to
  726. 27:40buy party masks—like a Donald Trump
  727. 27:43mask or one of then-Prime Minister Abe
  728. 27:46—and create driver's license images
  729. 27:48wearing those masks. Of course, these
  730. 27:52were documents made solely for testing
  731. 27:55purposes. We printed them out, verified
  732. 27:57they passed the thickness check, and
  733. 27:59when we followed the instructions to
  734. 28:01turn right while wearing the mask, it
  735. 28:03actually cleared the verification
  736. 28:05process.
  737. 28:06>> Wow. Granted, this was back in 2018 or
  738. 28:102019, and while companies have likely
  739. 28:13improved their technical accuracy
  740. 28:15within legal frameworks since then,
  741. 28:17generative AI and deepfakes can now
  742. 28:19create much more precise facial images
  743. 28:22than those masks. I believe today's
  744. 28:25methods have their limits. Oh.
  745. 28:27>> Ah,
  746. 28:28>> Yeah.
  747. 28:29>> And at a time like this, Mr. Takezaki
  748. 28:31...
  749. 28:31>> Yes. Our time has already flown by. I'm
  750. 28:34sorry. I learned so much. No, thank you
  751. 28:37as well.
  752. 28:38>> Please, keep asking him plenty of
  753. 28:39questions.
  754. 28:40>> I have to head to another job, sorry.
  755. 28:42Excuse me.
  756. 28:43>> Good work today.
  757. 28:46>> I'll just move this chair.
  758. 28:48>> I recall these discussions heating up
  759. 28:51in the U.S., and I feel like the OpenAI
  760. 28:53and Hugging Face situation was really
  761. 28:56the trigger for it.
  762. 28:57>> And that, well, was an example of AI
  763. 29:00running wild, wasn't it? Should we view
  764. 29:03this current incident as strictly the
  765. 29:05work of malicious individuals?
  766. 29:08>> Or can we assume there is zero
  767. 29:10possibility that an AI went rogue and
  768. 29:12orchestrated this? Or should we
  769. 29:15consider that there might be a slight
  770. 29:17chance of that? What do you think?
  771. 29:19>> This is an extremely good and yet
  772. 29:22extremely difficult question, so I will
  773. 29:25answer it carefully. I think that as
  774. 29:28long as we, the Japanese and Americans,
  775. 29:31are using well-known and widely used AI
  776. 29:34like OpenAI or Anthropic, it is
  777. 29:36unlikely that this kind of malicious
  778. 29:38runaway behavior will happen. This is
  779. 29:42because, as I mentioned earlier, the
  780. 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.