[
  {
    "start": 0.1,
    "end": 3.78,
    "text": "A disturbing,\nyet not really surprising article or"
  },
  {
    "start": 3.82,
    "end": 7.28,
    "text": "post was published by Anthropic yesterday."
  },
  {
    "start": 7.36,
    "end": 11.08,
    "text": "An article about an AI-orchestrated"
  },
  {
    "start": 11.24,
    "end": 14.76,
    "text": "cyber espionage campaign, a cyberattack"
  },
  {
    "start": 15.06,
    "end": 17.72,
    "text": "carried out with help of AI,\nwith help of Claude code."
  },
  {
    "start": 17.74,
    "end": 21.24,
    "text": "And it's an interesting article.\nI already read it"
  },
  {
    "start": 21.28,
    "end": 25.12,
    "text": "together with you here. So Anthropic,\nin this article,"
  },
  {
    "start": 25.16,
    "end": 28.84,
    "text": "describes a, a cyberattack on"
  },
  {
    "start": 28.9,
    "end": 32.74,
    "text": "various companies that\nwas carried out almost entirely with"
  },
  {
    "start": 32.759,
    "end": 36.74,
    "text": "help of AI,\nalmost entirely with help of Claude code,"
  },
  {
    "start": 36.78,
    "end": 40.58,
    "text": "by jailbreaking Claude code,\nby getting it to do"
  },
  {
    "start": 40.88,
    "end": 44.61,
    "text": "things it normally shouldn't do.\nAnd let's take a closer look at"
  },
  {
    "start": 44.68,
    "end": 48.6,
    "text": "how that played out.\n\"In mid-September 2025,"
  },
  {
    "start": 48.62,
    "end": 52.51,
    "text": "we, Anthropic,\ndetected suspicious activity that later,"
  },
  {
    "start": 52.54,
    "end": 56.5,
    "text": "investigation determined to be a highly\nsophisticated espionage campaign."
  },
  {
    "start": 56.54,
    "end": 59.98,
    "text": "The attackers used AI's agentic\ncapabilities to an"
  },
  {
    "start": 60.06,
    "end": 64.04,
    "text": "unprecedented degree,\nusing AI not just as an advisor, but to"
  },
  {
    "start": 64.12,
    "end": 67.83,
    "text": "execute the cyberattacks themselves.\"\nAnd that's, that"
  },
  {
    "start": 67.9,
    "end": 71.26,
    "text": "is huge that you can"
  },
  {
    "start": 71.36,
    "end": 74.9,
    "text": "already use AI, AI that is"
  },
  {
    "start": 74.92,
    "end": 78.72,
    "text": "published by Anthropic,\nAI models by Anthropic, and the"
  },
  {
    "start": 78.78,
    "end": 82.54,
    "text": "Claude code tool by Anthropic to carry out"
  },
  {
    "start": 82.6,
    "end": 85.04,
    "text": "cyberattacks. And, uh, here's why that is"
  },
  {
    "start": 85.06,
    "end": 89.04,
    "text": "important.\nToday we're living in a world where"
  },
  {
    "start": 89.1,
    "end": 92.54,
    "text": "the most capable models are the"
  },
  {
    "start": 92.6,
    "end": 95.88,
    "text": "models by OpenAI, Anthropic, X,"
  },
  {
    "start": 95.94,
    "end": 99.86,
    "text": "Google,\nand of course we've got some very capable"
  },
  {
    "start": 99.92,
    "end": 103.69,
    "text": "too,\nbut we're still at a point in time on a,"
  },
  {
    "start": 103.76,
    "end": 107.38,
    "text": "timeline, we are still at a point in time"
  },
  {
    "start": 107.42,
    "end": 110.96,
    "text": "where most of these models and the,\nthe most capable"
  },
  {
    "start": 111.02,
    "end": 114.62,
    "text": "models are controlled by companies,"
  },
  {
    "start": 115.12,
    "end": 117.84,
    "text": "most of them by companies in democracies."
  },
  {
    "start": 117.9,
    "end": 121.88,
    "text": "Now, I'll not say that these companies\nare saints or"
  },
  {
    "start": 122.28,
    "end": 126.12,
    "text": "don't do bad stuff, but they're not bad"
  },
  {
    "start": 126.24,
    "end": 129.56,
    "text": "actors in the sense of this article."
  },
  {
    "start": 129.58,
    "end": 132.59,
    "text": "They're not cyber, um, attackers"
  },
  {
    "start": 132.84,
    "end": 136.46,
    "text": "obviously. However, in the future,\nin the not too"
  },
  {
    "start": 136.6,
    "end": 139.84,
    "text": "distant future, we will be at a place"
  },
  {
    "start": 139.9,
    "end": 143.41,
    "text": "where these models,\nthese very capable models"
  },
  {
    "start": 144.02,
    "end": 147.75,
    "text": "will also be owned by bad actors"
  },
  {
    "start": 147.78,
    "end": 151.6,
    "text": "themselves.\nSo right now here in this article,"
  },
  {
    "start": 151.68,
    "end": 155.2,
    "text": "about a cyberattack that\nwas carried out with help of"
  },
  {
    "start": 155.28,
    "end": 159.0,
    "text": "Anthropic and Claude code.\nAnd it's bad enough that this is possible,"
  },
  {
    "start": 159.16,
    "end": 163.06,
    "text": "trick those tools\nand models into doing stuff they shouldn't"
  },
  {
    "start": 163.1,
    "end": 166.72,
    "text": "we'll get back to how that happened,\nof course, but that is"
  },
  {
    "start": 166.76,
    "end": 170.56,
    "text": "today. We still have to apply tricks,"
  },
  {
    "start": 170.64,
    "end": 173.86,
    "text": "uh, to abuse these models. In a"
  },
  {
    "start": 173.94,
    "end": 177.48,
    "text": "future,\nthese models will simply belong to the bad"
  },
  {
    "start": 177.66,
    "end": 181.42,
    "text": "actors themselves.\nThere will be open models that are capable"
  },
  {
    "start": 181.48,
    "end": 185.35,
    "text": "enough of doing that,\nso then certain control mechanisms,"
  },
  {
    "start": 185.4,
    "end": 189.14,
    "text": "get back here in this article,\nwon't even be there anymore."
  },
  {
    "start": 189.2,
    "end": 193.0,
    "text": "We'll be in a future where bad actors have"
  },
  {
    "start": 193.04,
    "end": 196.97,
    "text": "direct,\nuncontrolled access to very capable models"
  },
  {
    "start": 197.14,
    "end": 200.94,
    "text": "can be fine-tuned for their purposes,\nthat can be"
  },
  {
    "start": 201.04,
    "end": 204.9,
    "text": "trained for their purposes,\nthat can use tools that were"
  },
  {
    "start": 205.0,
    "end": 207.5,
    "text": "purpose built for malicious stuff."
  },
  {
    "start": 207.58,
    "end": 211.38,
    "text": "That is where we're heading to,\nand even today where we're not there"
  },
  {
    "start": 211.52,
    "end": 215.04,
    "text": "yet or where this is still a niche,\neven today, we"
  },
  {
    "start": 215.2,
    "end": 219.19,
    "text": "have fully or almost fully AI-controlled,"
  },
  {
    "start": 219.42,
    "end": 223.3,
    "text": "uh, cyberattacks. So this\nis definitely a scary article"
  },
  {
    "start": 223.36,
    "end": 227.02,
    "text": "and, and a scary future,\nwhich makes it very clear that"
  },
  {
    "start": 227.12,
    "end": 229.56,
    "text": "cybersecurity and"
  },
  {
    "start": 229.62,
    "end": 233.4,
    "text": "preventing attacks will be a super"
  },
  {
    "start": 233.52,
    "end": 237.42,
    "text": "big challenge. It always has been,\nbut now with AI where everything can be"
  },
  {
    "start": 237.52,
    "end": 240.84,
    "text": "quicker and more automated\nand harder to trace back to"
  },
  {
    "start": 240.88,
    "end": 244.64,
    "text": "individuals,\nit will be an even more important"
  },
  {
    "start": 244.66,
    "end": 247.84,
    "text": "topic. But back to this article.\n\"The threat"
  },
  {
    "start": 247.94,
    "end": 251.76,
    "text": "actor,\nwhom we assess with high confidence"
  },
  {
    "start": 251.98,
    "end": 255.56,
    "text": "state-sponsored group,\nmanipulated our Claude code"
  },
  {
    "start": 255.98,
    "end": 259.6,
    "text": "tool into attempting infiltration into\nroughly 30"
  },
  {
    "start": 259.64,
    "end": 263.63,
    "text": "global targets and succeeded in a small\nnumber of cases.\" So it was not just"
  },
  {
    "start": 263.66,
    "end": 267.5,
    "text": "an attempt. They succeeded. \"The operation"
  },
  {
    "start": 267.54,
    "end": 271.12,
    "text": "targeted large tech companies,\nfinancial institutions,"
  },
  {
    "start": 271.16,
    "end": 273.81,
    "text": "companies, and government agencies."
  },
  {
    "start": 274.02,
    "end": 277.58,
    "text": "Uh, we believe this\nis the first documented case of a"
  },
  {
    "start": 277.64,
    "end": 281.42,
    "text": "cyberattack executed without substantial\nhuman intervention.\" This is so"
  },
  {
    "start": 281.52,
    "end": 283.86,
    "text": "big, without substantial human"
  },
  {
    "start": 283.92,
    "end": 287.69,
    "text": "intervention. And again, that\nis why this is"
  },
  {
    "start": 287.74,
    "end": 291.54,
    "text": "such a scary future where bad actors don't\neven"
  },
  {
    "start": 291.6,
    "end": 295.26,
    "text": "need to rely on, um, these controlled"
  },
  {
    "start": 295.52,
    "end": 298.74,
    "text": "AI models, uh,\nwhich they still have to rely on today."
  },
  {
    "start": 298.75,
    "end": 301.969,
    "text": "\"Upon detecting this activity,\nwe immediately launched an investigation"
  },
  {
    "start": 302.02,
    "end": 306.0,
    "text": "understand its scope and nature. Uh,\nover the following 10 days, as we"
  },
  {
    "start": 306.02,
    "end": 309.96,
    "text": "mapped the severity\nand full extent of the operation,"
  },
  {
    "start": 309.98,
    "end": 313.84,
    "text": "they were identified.\" So again,\nthese were really regular"
  },
  {
    "start": 313.94,
    "end": 316.84,
    "text": "Claude accounts. These people were using"
  },
  {
    "start": 316.9,
    "end": 319.91,
    "text": "Claude, the model hosted by Anthropic."
  },
  {
    "start": 319.92,
    "end": 323.64,
    "text": "They did not kind of steal it,\nrun it on their own servers"
  },
  {
    "start": 323.74,
    "end": 327.56,
    "text": "models.\nThey used the models you can use too via"
  },
  {
    "start": 327.6,
    "end": 331.42,
    "text": "via Claude code.\nThis campaign has substantial"
  },
  {
    "start": 331.44,
    "end": 335.28,
    "text": "implications for cybersecurity in the age\nof AI agents, as I just said,"
  },
  {
    "start": 335.34,
    "end": 339.29,
    "text": "because everything can be automated\nand it's already possible today, uh,"
  },
  {
    "start": 339.34,
    "end": 342.09,
    "text": "where there are guardrails in place."
  },
  {
    "start": 342.1,
    "end": 345.8,
    "text": "And again, think of\nthat future where we have no guardrails"
  },
  {
    "start": 345.92,
    "end": 349.88,
    "text": "actors don't have guardrails. \"Uh,\nthese attacks are likely"
  },
  {
    "start": 349.92,
    "end": 351.9,
    "text": "to only grow in their effectiveness."
  },
  {
    "start": 351.94,
    "end": 355.68,
    "text": "To keep pace with this rapidly advancing\nthreat, we have expanded our"
  },
  {
    "start": 355.7,
    "end": 359.65,
    "text": "detection capabilities\nand developed better classifiers to flag"
  },
  {
    "start": 359.72,
    "end": 363.472,
    "text": "activities.\" And that's the important part\nhere.This is what"
  },
  {
    "start": 363.482,
    "end": 366.832,
    "text": "Anthropic is trying to do today to make\nsure that their"
  },
  {
    "start": 366.872,
    "end": 370.742,
    "text": "models can't be abused for malicious\ntasks,"
  },
  {
    "start": 370.812,
    "end": 374.612,
    "text": "that Claude code can't be abused\nand ultimately,"
  },
  {
    "start": 374.632,
    "end": 377.952,
    "text": "their APIs,\nwhich Claude code uses in the end."
  },
  {
    "start": 378.012,
    "end": 381.972,
    "text": "This all won't matter at all in the future\nbecause"
  },
  {
    "start": 382.052,
    "end": 385.912,
    "text": "in a future where bad actors themselves\nhave their own models"
  },
  {
    "start": 385.972,
    "end": 389.762,
    "text": "running on their own servers,\nthese guardrails won't"
  },
  {
    "start": 389.792,
    "end": 391.902,
    "text": "matter. Well,\nobviously they will still matter."
  },
  {
    "start": 391.972,
    "end": 395.552,
    "text": "Obviously,\nyou still want to make it easier than it's"
  },
  {
    "start": 395.632,
    "end": 399.072,
    "text": "obviously not all bad actors will have\ntheir"
  },
  {
    "start": 399.152,
    "end": 402.912,
    "text": "own malicious models, but especially\nif we're talking about"
  },
  {
    "start": 402.932,
    "end": 406.512,
    "text": "state-controlled bad actors or big"
  },
  {
    "start": 406.572,
    "end": 410.352,
    "text": "cyber attacker groups. Let's, let's be"
  },
  {
    "start": 410.372,
    "end": 414.252,
    "text": "real.\nOf course they will have access to their"
  },
  {
    "start": 414.292,
    "end": 417.882,
    "text": "on their own servers,\nso this will not matter at"
  },
  {
    "start": 417.992,
    "end": 421.932,
    "text": "all in that future. Obviously,\nit will matter still because you don't"
  },
  {
    "start": 421.972,
    "end": 425.832,
    "text": "make it easier and it will at least filter\nout a significant"
  },
  {
    "start": 425.932,
    "end": 429.812,
    "text": "group of potential bad actors\nthat don't have access to their own"
  },
  {
    "start": 429.872,
    "end": 433.852,
    "text": "models. So yeah, it's important,\nbut it will not be enough."
  },
  {
    "start": 433.912,
    "end": 437.712,
    "text": "Companies themselves need to ramp up"
  },
  {
    "start": 437.772,
    "end": 441.732,
    "text": "their cybersecurity game, which\nis easier said than done because"
  },
  {
    "start": 441.772,
    "end": 445.752,
    "text": "it has been true for the last 10 years,\nof course, even without AI, but it's"
  },
  {
    "start": 445.772,
    "end": 448.872,
    "text": "becoming even more important in the age of\nAI."
  },
  {
    "start": 450.132,
    "end": 453.612,
    "text": "So yeah, that is, that\nis the big problem here."
  },
  {
    "start": 453.652,
    "end": 457.032,
    "text": "Now, let's see how the, uh,\ncyberattack worked."
  },
  {
    "start": 457.132,
    "end": 460.762,
    "text": "Uh,\nthe attack relied on several features of"
  },
  {
    "start": 460.792,
    "end": 464.432,
    "text": "exist or were in much more nascent form\njust a year ago."
  },
  {
    "start": 464.452,
    "end": 468.172,
    "text": "Intelligence:\nModels' general levels of capability have"
  },
  {
    "start": 468.272,
    "end": 472.152,
    "text": "point that they can follow complicance-\ncomplex instructions and understand"
  },
  {
    "start": 472.192,
    "end": 476.012,
    "text": "context in ways that make very\nsophisticated tasks possible."
  },
  {
    "start": 476.052,
    "end": 479.722,
    "text": "Not only that,\nbut several of their well-developed"
  },
  {
    "start": 479.752,
    "end": 483.212,
    "text": "particular software coding,\nlend themselves to being used in"
  },
  {
    "start": 483.252,
    "end": 487.122,
    "text": "Sure, because now with models that\nare smarter,"
  },
  {
    "start": 487.152,
    "end": 490.782,
    "text": "and just to be very clear here,\nwe're still talking about models that"
  },
  {
    "start": 490.892,
    "end": 494.872,
    "text": "just generate tokens,\nbut of course by generating these tokens"
  },
  {
    "start": 494.932,
    "end": 498.922,
    "text": "that are able, they\nare able to describe the usage of tools"
  },
  {
    "start": 498.952,
    "end": 502.452,
    "text": "tools, they become more capable.\nWith all the"
  },
  {
    "start": 502.472,
    "end": 506.092,
    "text": "fine-tuning they received,\nthey also generate more tokens that are"
  },
  {
    "start": 506.352,
    "end": 510.332,
    "text": "likely to be or more likely to be the\ntokens you want to generate, so that"
  },
  {
    "start": 510.372,
    "end": 514.292,
    "text": "is what intelligence means here.\nThey're not really intelligent, but they"
  },
  {
    "start": 514.352,
    "end": 518.182,
    "text": "have been tuned especially for software\ndevelopment such that"
  },
  {
    "start": 518.212,
    "end": 521.962,
    "text": "they are much more likely to generate\nmeaningful output"
  },
  {
    "start": 522.272,
    "end": 526.092,
    "text": "and especially also output\nthat allows them to describe tool"
  },
  {
    "start": 526.172,
    "end": 529.892,
    "text": "use and then use those tools,\nso execute code"
  },
  {
    "start": 529.902,
    "end": 533.732,
    "text": "that does something. Uh, for example,\nsend an HTTP request and so"
  },
  {
    "start": 533.772,
    "end": 537.572,
    "text": "on, and it's that combination\nthat makes them more capable"
  },
  {
    "start": 537.912,
    "end": 541.112,
    "text": "in the end. And of course, yeah,\nthat's exactly what you need for automated"
  },
  {
    "start": 541.132,
    "end": 545.012,
    "text": "cyberattacks,\nbecause you need a model that's able to"
  },
  {
    "start": 545.072,
    "end": 547.272,
    "text": "follow your instructions related to that."
  },
  {
    "start": 547.312,
    "end": 550.912,
    "text": "You need a model that's able to send HTTP\nrequests, phishing"
  },
  {
    "start": 550.972,
    "end": 554.662,
    "text": "emails, whatever, and all that\nis what these models can do"
  },
  {
    "start": 554.752,
    "end": 558.242,
    "text": "quite well in the end. Agency:\nModels can act as agents."
  },
  {
    "start": 558.272,
    "end": 561.872,
    "text": "That is,\nthey can run in loops where they take"
  },
  {
    "start": 561.932,
    "end": 564.812,
    "text": "tasks and make decisions with only minimal\noccasional human"
  },
  {
    "start": 564.872,
    "end": 567.432,
    "text": "input. That's another important step."
  },
  {
    "start": 567.512,
    "end": 571.242,
    "text": "This is what allows these, um, systems\nor these"
  },
  {
    "start": 571.272,
    "end": 575.192,
    "text": "attacks here in this case to work with\nonly minimal human input"
  },
  {
    "start": 576.132,
    "end": 580.032,
    "text": "because these models and the software\nthat uses"
  },
  {
    "start": 580.072,
    "end": 583.712,
    "text": "these models can go for so much longer\nand it's important to"
  },
  {
    "start": 583.752,
    "end": 587.672,
    "text": "differentiate here. The model is,\nis still just the"
  },
  {
    "start": 587.712,
    "end": 591.332,
    "text": "thing that receives a prompt\nand sends back some tokens."
  },
  {
    "start": 591.352,
    "end": 595.292,
    "text": "That, that has not changed.\nBut it's the software, Claude Code, for"
  },
  {
    "start": 595.372,
    "end": 599.252,
    "text": "example, that then takes that output\nand sends back another"
  },
  {
    "start": 599.312,
    "end": 602.742,
    "text": "message to the same API with that output,\nwith the original"
  },
  {
    "start": 602.852,
    "end": 606.572,
    "text": "task, with some meta instructions like,\n\"Please check if that"
  },
  {
    "start": 606.612,
    "end": 609.412,
    "text": "output answers the question by the user."
  },
  {
    "start": 609.472,
    "end": 612.932,
    "text": "You got these tools available,\nplease tell me if you want to use a tool.\""
  },
  {
    "start": 612.992,
    "end": 616.912,
    "text": "how the software around the models in the\nend makes these"
  },
  {
    "start": 616.992,
    "end": 620.711,
    "text": "models more capable,\nnot because the model does everything on"
  },
  {
    "start": 620.752,
    "end": 624.352,
    "text": "the model is capable of giving the\nsoftware the"
  },
  {
    "start": 624.392,
    "end": 628.372,
    "text": "result it needs.\nThe software then feeds these enriched"
  },
  {
    "start": 628.412,
    "end": 632.112,
    "text": "the model and it's this loop\nthat keeps the whole system going and that"
  },
  {
    "start": 632.192,
    "end": 635.972,
    "text": "leads to these agentic systems that can go"
  },
  {
    "start": 636.052,
    "end": 639.752,
    "text": "on for longer, that can use tools and\nthat require less"
  },
  {
    "start": 639.812,
    "end": 643.692,
    "text": "human input. And yeah, tools, that\nis therefore the other missing piece"
  },
  {
    "start": 643.732,
    "end": 646.972,
    "text": "here, of course,\nthat models have access to a wide array of"
  },
  {
    "start": 647.012,
    "end": 650.612,
    "text": "tools. They can now search the web,\nretrieve data, perform many other"
  },
  {
    "start": 650.692,
    "end": 653.662,
    "text": "actions that were previously the sole\ndomain of human operators."
  },
  {
    "start": 653.692,
    "end": 657.492,
    "text": "In the case of cyberattacks,\nthe tools might include password crackers,"
  },
  {
    "start": 657.552,
    "end": 660.392,
    "text": "scanners and other security-related\nsoftware."
  },
  {
    "start": 660.412,
    "end": 664.152,
    "text": "Because again,\nit's not all just GitHub MCPs."
  },
  {
    "start": 664.192,
    "end": 667.692,
    "text": "It can be all kind of tools you could, uh,\nexpose"
  },
  {
    "start": 667.752,
    "end": 671.572,
    "text": "to your, um, model or to the software\nthat uses"
  },
  {
    "start": 671.612,
    "end": 674.332,
    "text": "these models and\nthat runs these agentic tasks."
  },
  {
    "start": 674.412,
    "end": 678.152,
    "text": "So they got a nice diagram in this\narticle, but in the end"
  },
  {
    "start": 678.252,
    "end": 681.721,
    "text": "the attack played out relatively, uh,"
  },
  {
    "start": 681.771,
    "end": 685.122,
    "text": "simple. They,\nthey describe it in greater detail down"
  },
  {
    "start": 685.152,
    "end": 689.012,
    "text": "there. They convinced Claude Code to do"
  },
  {
    "start": 689.031,
    "end": 691.192,
    "text": "stuff it normally shouldn't be able to do."
  },
  {
    "start": 691.232,
    "end": 694.781,
    "text": "They had to convince Claude, which\nis extensively trained to avoid harmful"
  },
  {
    "start": 694.852,
    "end": 696.652,
    "text": "behaviors, to engage in the attack."
  },
  {
    "start": 696.672,
    "end": 700.652,
    "text": "They did so by jailbreaking it,\neffectively tricking it to bypass"
  },
  {
    "start": 700.672,
    "end": 704.452,
    "text": "its guardrails,\nand that's the part where Anthropic"
  },
  {
    "start": 704.472,
    "end": 708.332,
    "text": "back better, not just in Claude Code\nbut in the"
  },
  {
    "start": 708.452,
    "end": 712.292,
    "text": "models themselves,\nso on their API where they scan all the"
  },
  {
    "start": 712.332,
    "end": 716.142,
    "text": "requests that reach their models,\nso to say, and take"
  },
  {
    "start": 716.152,
    "end": 719.942,
    "text": "better... eh,\nthey try to do a better job at detecting"
  },
  {
    "start": 720.031,
    "end": 724.021,
    "text": "injections in the end because these\nattackers broke down their attacks into"
  },
  {
    "start": 724.112,
    "end": 727.962,
    "text": "small, seemingly innocent tasks\nthat Claude would execute without being"
  },
  {
    "start": 727.992,
    "end": 730.762,
    "text": "provided the full context of their\nmalicious purpose."
  },
  {
    "start": 730.762,
    "end": 734.712,
    "text": "They also told Claude that it\nwas an employee of a legitimate"
  },
  {
    "start": 734.732,
    "end": 738.632,
    "text": "cybersecurity firm and\nwas being used in defensive testing.That's"
  },
  {
    "start": 738.652,
    "end": 742.432,
    "text": "good old trick. I think that\nis how jailbreaking was already done two"
  },
  {
    "start": 742.492,
    "end": 745.722,
    "text": "years ago with the early ChatGPT models."
  },
  {
    "start": 745.812,
    "end": 749.692,
    "text": "Eh,\nyou tell it that you need this information"
  },
  {
    "start": 749.792,
    "end": 753.052,
    "text": "and it'll happily expose its system\nprompt."
  },
  {
    "start": 753.092,
    "end": 756.952,
    "text": "Kind of a simplification\nbut that's still how prompt injections can"
  },
  {
    "start": 756.972,
    "end": 760.802,
    "text": "days.\nThat you try to apply various techniques"
  },
  {
    "start": 760.832,
    "end": 764.692,
    "text": "are very interesting techniques\nwhen it comes to that,"
  },
  {
    "start": 764.972,
    "end": 768.232,
    "text": "eh,\nincluding the use of special tokens you"
  },
  {
    "start": 768.242,
    "end": 771.372,
    "text": "message to tr- to get the"
  },
  {
    "start": 771.472,
    "end": 774.822,
    "text": "model to generate output it normally\nshouldn't"
  },
  {
    "start": 774.872,
    "end": 777.972,
    "text": "generate.\nThe attackers then initiated the second"
  },
  {
    "start": 778.092,
    "end": 781.241,
    "text": "involved Claude code ins-\ninspecting the target organization's"
  },
  {
    "start": 781.272,
    "end": 784.692,
    "text": "So yeah,\nthat's then essentially what Claude code"
  },
  {
    "start": 784.732,
    "end": 788.652,
    "text": "did. Then with minimal human input, um, it"
  },
  {
    "start": 788.732,
    "end": 791.972,
    "text": "is in the end then used its agentic\ncapabilities, its"
  },
  {
    "start": 792.032,
    "end": 795.741,
    "text": "tools, to really, um,"
  },
  {
    "start": 795.832,
    "end": 799.812,
    "text": "scan networks, write code, and do all that"
  },
  {
    "start": 799.822,
    "end": 803.702,
    "text": "stuff without a human telling it exactly\nwhat to do."
  },
  {
    "start": 803.752,
    "end": 807.692,
    "text": "So, it was, as mentioned earlier, uh,\na fully or almost fully"
  },
  {
    "start": 807.752,
    "end": 811.292,
    "text": "automated attack. So,\nin the next phases of the attack, Claude"
  },
  {
    "start": 811.372,
    "end": 815.262,
    "text": "identified and tested security\nvulnerabilities in the target"
  },
  {
    "start": 815.352,
    "end": 818.392,
    "text": "systems by researching\nand writing its own exploit"
  },
  {
    "start": 818.512,
    "end": 822.252,
    "text": "code. Having done so, the framework\nwas able to use Claude to"
  },
  {
    "start": 822.292,
    "end": 825.952,
    "text": "harvest credentials, usernames\nand passwords that allowed it to further"
  },
  {
    "start": 825.972,
    "end": 828.192,
    "text": "then extract a large amount of private\ndata."
  },
  {
    "start": 828.232,
    "end": 832.192,
    "text": "So, it did really research,\nwrite the code to"
  },
  {
    "start": 832.252,
    "end": 835.982,
    "text": "get into systems, uh,\nof other companies and then in those"
  },
  {
    "start": 836.092,
    "end": 839.752,
    "text": "systems write more code to extract data,"
  },
  {
    "start": 839.852,
    "end": 843.362,
    "text": "um, and- and-\nand compromise these systems and- and- and"
  },
  {
    "start": 843.692,
    "end": 847.062,
    "text": "do bad stuff in there once it\nwas in there."
  },
  {
    "start": 847.092,
    "end": 850.952,
    "text": "The highest privileged accounts\nwere identified, backdoors"
  },
  {
    "start": 851.012,
    "end": 854.752,
    "text": "all the stuff that happened after it\nwas in the systems and data were exfil-"
  },
  {
    "start": 854.992,
    "end": 858.292,
    "text": "exfiltrated with minimal human\nsupervision."
  },
  {
    "start": 858.312,
    "end": 861.692,
    "text": "In a final phase,\nthe attackers had Claude produce"
  },
  {
    "start": 861.732,
    "end": 864.992,
    "text": "the attack,\ncreating helpful files with the stolen"
  },
  {
    "start": 865.112,
    "end": 868.712,
    "text": "analyzed which would assist the framework\nin planning the next stage of the"
  },
  {
    "start": 868.772,
    "end": 872.732,
    "text": "threats actor, eh,\nof the threat actor's cyber operations."
  },
  {
    "start": 872.792,
    "end": 876.412,
    "text": "Overall, the threat actor\nwas able to use AI to perform 80 to 90% of"
  },
  {
    "start": 876.432,
    "end": 880.292,
    "text": "campaign with human intervention required\nonly sporadically, perhaps four to"
  },
  {
    "start": 880.312,
    "end": 882.942,
    "text": "six critical decision points per hacking\ncampaign."
  },
  {
    "start": 882.992,
    "end": 886.952,
    "text": "That is nothing. That is nothing. That\nis such"
  },
  {
    "start": 887.032,
    "end": 890.652,
    "text": "a scale at which you can run these attacks\nand again, especially in a"
  },
  {
    "start": 890.732,
    "end": 894.552,
    "text": "future where you don't have to work\nagainst certain guardrails, where you can"
  },
  {
    "start": 894.712,
    "end": 898.552,
    "text": "just focus on getting the job done\nand you have to spend"
  },
  {
    "start": 898.612,
    "end": 900.462,
    "text": "energy on getting around guardrails."
  },
  {
    "start": 900.512,
    "end": 903.912,
    "text": "That is really a scary future.\nThis degree of"
  },
  {
    "start": 903.992,
    "end": 907.892,
    "text": "automation is really, really, uh, scary"
  },
  {
    "start": 907.902,
    "end": 909.902,
    "text": "here. Claude didn't always work perfectly."
  },
  {
    "start": 909.912,
    "end": 913.372,
    "text": "It occasionally hallucinated credentials\nor claimed to have extracted secret"
  },
  {
    "start": 913.392,
    "end": 915.742,
    "text": "information that was in fact publicly\navailable."
  },
  {
    "start": 915.792,
    "end": 919.782,
    "text": "This remains an obstacle to fully, uh,\nautonomous cyber attacks and this is"
  },
  {
    "start": 919.832,
    "end": 922.912,
    "text": "not just an obstacle for cyber attacks,\nthis is of course an obstacle for"
  },
  {
    "start": 922.972,
    "end": 926.752,
    "text": "everybody,\nfor us developers too because"
  },
  {
    "start": 926.812,
    "end": 930.672,
    "text": "problem and will stay a problem because as\nI mentioned before, it's so"
  },
  {
    "start": 930.732,
    "end": 933.832,
    "text": "easy to forget but these\nare token generation"
  },
  {
    "start": 934.032,
    "end": 936.981,
    "text": "machines. Always have been,\nalways will be."
  },
  {
    "start": 936.992,
    "end": 938.832,
    "text": "The large language models, I mean."
  },
  {
    "start": 938.872,
    "end": 942.772,
    "text": "They are generating tokens and they\nare generating"
  },
  {
    "start": 942.832,
    "end": 946.772,
    "text": "the most likely token as the next token\nbased on all the"
  },
  {
    "start": 946.782,
    "end": 950.172,
    "text": "tokens that came before it, and that\nis something that can and"
  },
  {
    "start": 950.272,
    "end": 953.232,
    "text": "always will have the danger of"
  },
  {
    "start": 953.252,
    "end": 956.832,
    "text": "hallucinating. So, that of course\nis a problem in"
  },
  {
    "start": 956.872,
    "end": 960.752,
    "text": "general. Good to see\nthat it can then also be helpful"
  },
  {
    "start": 960.832,
    "end": 964.532,
    "text": "when it comes to defending against\nmalicious tasks because"
  },
  {
    "start": 964.572,
    "end": 968.322,
    "text": "those also are hurt by hallucination\nbut of course that is a general"
  },
  {
    "start": 968.322,
    "end": 971.882,
    "text": "problem, uh, we face, uh,\nand it will not be a"
  },
  {
    "start": 971.912,
    "end": 974.602,
    "text": "significant, uh, defense mechanism"
  },
  {
    "start": 974.692,
    "end": 978.652,
    "text": "unfortunately. Because in the end\nif everything's automated, it's"
  },
  {
    "start": 978.732,
    "end": 982.322,
    "text": "just a question of scale and\nif some attacks fail because of"
  },
  {
    "start": 982.352,
    "end": 986.242,
    "text": "hallucination, well, does\nthat really matter if you can run"
  },
  {
    "start": 986.252,
    "end": 988.462,
    "text": "thousands of attacks in parallel?"
  },
  {
    "start": 988.532,
    "end": 991.682,
    "text": "I'm not sure it does.\nCybersecurity implications."
  },
  {
    "start": 991.712,
    "end": 995.492,
    "text": "The barriers to performing sophisticated\ncyber attacks have dropped"
  },
  {
    "start": 995.512,
    "end": 998.731,
    "text": "substantially and we predict\nthat they'll continue to do so."
  },
  {
    "start": 998.752,
    "end": 1002.492,
    "text": "With the correct setup,\nthreat actors can now use agentic AI"
  },
  {
    "start": 1002.512,
    "end": 1006.312,
    "text": "extended periods to do the work of entire\nteams of experienced hackers."
  },
  {
    "start": 1006.322,
    "end": 1010.152,
    "text": "So kind of the same thing\nthat applies to normal software"
  },
  {
    "start": 1010.212,
    "end": 1013.892,
    "text": "also be more productive with AI,\nand I got a video coming up on"
  },
  {
    "start": 1013.932,
    "end": 1017.632,
    "text": "way, just that that\nis the case for malicious tasks,"
  },
  {
    "start": 1017.692,
    "end": 1021.412,
    "text": "even worse because there you don't even\nhave to"
  },
  {
    "start": 1021.512,
    "end": 1024.642,
    "text": "care about things like code quality."
  },
  {
    "start": 1024.672,
    "end": 1028.082,
    "text": "Obviously you want to have a successful\nattack but in the end if everything's"
  },
  {
    "start": 1028.172,
    "end": 1031.732,
    "text": "automated,\nyou also can care a lot about just the"
  },
  {
    "start": 1031.832,
    "end": 1035.672,
    "text": "scale.\nAnd if you can automate thousands of"
  },
  {
    "start": 1035.732,
    "end": 1039.472,
    "text": "attacks to run in parallel,\nit doesn't really"
  },
  {
    "start": 1039.532,
    "end": 1043.512,
    "text": "matter to you if you might have code\nquality problems"
  },
  {
    "start": 1043.552,
    "end": 1047.132,
    "text": "or anything like that. So, uh,"
  },
  {
    "start": 1047.212,
    "end": 1051.132,
    "text": "already with the systems today where you\ncould argue about potential"
  },
  {
    "start": 1051.252,
    "end": 1055.182,
    "text": "problems they have\nwhen it comes to generating code,"
  },
  {
    "start": 1055.252,
    "end": 1058.682,
    "text": "matter for attacks like this because you\nneed a result that's just good"
  },
  {
    "start": 1058.732,
    "end": 1062.672,
    "text": "enough. And again, chances\nare definitely high that results"
  },
  {
    "start": 1062.712,
    "end": 1066.652,
    "text": "will also get better in the future\nand we'll be dealing with systems"
  },
  {
    "start": 1066.662,
    "end": 1070.232,
    "text": "that don't even have guardrails.\nAnd as they say here, less"
  },
  {
    "start": 1070.272,
    "end": 1073.712,
    "text": "experienced and resourced groups can now\npotentially perform"
  },
  {
    "start": 1073.732,
    "end": 1076.932,
    "text": "large-scale, uh, attacks of this nature."
  },
  {
    "start": 1077.002,
    "end": 1080.652,
    "text": "This attack is an escalation even on the\nwipe hacking findings we"
  },
  {
    "start": 1080.692,
    "end": 1084.621,
    "text": "reported this summer. In those operations,\nhumans were much, uh, still in"
  },
  {
    "start": 1084.652,
    "end": 1087.572,
    "text": "the loop, uh, directing the operations."
  },
  {
    "start": 1087.652,
    "end": 1089.492,
    "text": "Here, human involvement was much less"
  },
  {
    "start": 1089.552,
    "end": 1093.032,
    "text": "frequent. And although we have"
  },
  {
    "start": 1093.052,
    "end": 1096.972,
    "text": "visibility into Claude usage,\nthis case study probably reflects"
  },
  {
    "start": 1097.012,
    "end": 1100.892,
    "text": "of behavior across frontier AI models\nand demonstrates how threat actors"
  },
  {
    "start": 1100.932,
    "end": 1104.702,
    "text": "are adapting their operations. By the way,\nthis is one case that was"
  },
  {
    "start": 1104.752,
    "end": 1108.625,
    "text": "caught by Anthropic.... not sure\nif all the cases are being caught, also"
  },
  {
    "start": 1108.716,
    "end": 1112.656,
    "text": "by Google, um, OpenAI and so on.\nThis raises"
  },
  {
    "start": 1112.676,
    "end": 1116.176,
    "text": "an important question.\nIf AI models can be misused for"
  },
  {
    "start": 1116.186,
    "end": 1119.146,
    "text": "scale, why continue to develop\nand release them?"
  },
  {
    "start": 1119.156,
    "end": 1123.136,
    "text": "The answer is that very, that very, a-\nabilities that allow Claude to be used in"
  },
  {
    "start": 1123.176,
    "end": 1126.005,
    "text": "these attacks also make it crucial for\ncyber defense."
  },
  {
    "start": 1126.035,
    "end": 1129.466,
    "text": "Well, (smacks lips) uh,\nthat's kind of a weak argument, I'll say,"
  },
  {
    "start": 1129.495,
    "end": 1133.406,
    "text": "because if you have one thing\nthat makes a problem much"
  },
  {
    "start": 1133.436,
    "end": 1137.326,
    "text": "bigger, saying, \"Yeah,\nbut it can also help with the solution,\""
  },
  {
    "start": 1137.376,
    "end": 1141.116,
    "text": "kind of bad, right? So, uh, I-\nI'm not really"
  },
  {
    "start": 1141.176,
    "end": 1144.505,
    "text": "sure about that.\nIf we would not have these models..."
  },
  {
    "start": 1144.535,
    "end": 1147.636,
    "text": "And just to be clear, that\nis not (laughs) something that's going to"
  },
  {
    "start": 1147.656,
    "end": 1151.636,
    "text": "But if we would not have them,\nit would probably be better than"
  },
  {
    "start": 1151.716,
    "end": 1155.636,
    "text": "in the context of cyberattacks\nand defense because"
  },
  {
    "start": 1155.645,
    "end": 1158.396,
    "text": "will always be one step, uh, behind."
  },
  {
    "start": 1158.476,
    "end": 1162.416,
    "text": "So, uh,\nI definitely see these tools more as an"
  },
  {
    "start": 1162.456,
    "end": 1166.055,
    "text": "advantage for the attackers\nand a big disadvantage, uh, for"
  },
  {
    "start": 1166.136,
    "end": 1169.436,
    "text": "the, uh, companies\nthat have to defend against these attacks."
  },
  {
    "start": 1169.456,
    "end": 1173.356,
    "text": "So that's kind of a weak argument,\nmy argument would be, it doesn't"
  },
  {
    "start": 1173.476,
    "end": 1177.015,
    "text": "matter if Anthropic, OpenAI and so on"
  },
  {
    "start": 1177.055,
    "end": 1180.696,
    "text": "continue developing AI models,\nand obviously they will, just to be very"
  },
  {
    "start": 1180.755,
    "end": 1184.596,
    "text": "clear.\nAnd there are way more arguments to be"
  },
  {
    "start": 1184.666,
    "end": 1188.275,
    "text": "cybersecurity. This is just one very i-\nimportant and problematic"
  },
  {
    "start": 1188.336,
    "end": 1192.035,
    "text": "field, but there are tons of discussions,\nincluding philosophical"
  },
  {
    "start": 1192.096,
    "end": 1196.055,
    "text": "discussions we could have about AI and\nif it's good that it's there or"
  },
  {
    "start": 1196.116,
    "end": 1199.396,
    "text": "not, but they all don't matter. It\nis there, it will stay there,"
  },
  {
    "start": 1199.436,
    "end": 1202.416,
    "text": "these companies will continuing to develop\nthese models."
  },
  {
    "start": 1202.436,
    "end": 1205.456,
    "text": "And even if they wouldn't,\nthe technology is there."
  },
  {
    "start": 1205.495,
    "end": 1209.356,
    "text": "Bad actors will have access to their own\nmodels in the"
  },
  {
    "start": 1209.396,
    "end": 1213.136,
    "text": "future. It does not matter at all\nif companies like Anthropic or"
  },
  {
    "start": 1213.235,
    "end": 1217.055,
    "text": "OpenAI continue. The technology\nis there and the"
  },
  {
    "start": 1217.096,
    "end": 1220.456,
    "text": "problems with it are also there,\ntherefore, and they will stay"
  },
  {
    "start": 1220.515,
    "end": 1224.495,
    "text": "here. That would be my argument.\nThis argument here doesn't make"
  },
  {
    "start": 1224.515,
    "end": 1228.285,
    "text": "too much sense to me. And therefore,\ndefinitely scary."
  },
  {
    "start": 1228.376,
    "end": 1231.696,
    "text": "A scary world also from a cybersecurity\nperspective."
  },
  {
    "start": 1231.795,
    "end": 1235.636,
    "text": "As I mentioned, I only see that, uh,\nbecoming worse"
  },
  {
    "start": 1235.755,
    "end": 1239.076,
    "text": "in the future, and therefore,\nmaybe a career in"
  },
  {
    "start": 1239.176,
    "end": 1242.815,
    "text": "cybersecurity is worth a second look\nbecause yeah, that will"
  },
  {
    "start": 1242.856,
    "end": 1244.156,
    "text": "be important."
  }
]