[
  {
    "start": 0.62,
    "end": 5.78,
    "text": "Have you ever been in a conversation with someone who just insists on their opinion no matter whether they're right or wrong?"
  },
  {
    "start": 6.4,
    "end": 11.4,
    "text": "It can be pretty frustrating and, In the group it can make fighting-the-right decision almost impossible."
  },
  {
    "start": 11.98,
    "end": 19.54,
    "text": "There goes your dinner plan because after twenty minutes of arguing everyone ends up settling for that pizza place around the corner That you already know is terrible."
  },
  {
    "start": 20.56,
    "end": 22.819,
    "text": "Now bad pizza has hardly at the end of the world."
  },
  {
    "start": 23.56,
    "end": 25.36,
    "text": "But what if this same kind of stubbornness shows up?"
  },
  {
    "start": 25.5,
    "end": 26.36,
    "text": "an AI system?"
  },
  {
    "start": 27.319,
    "end": 33.8,
    "text": "when multiple AIA agents are making decisions together One agent refusing to back down can have some pretty serious consequences."
  },
  {
    "start": 34.64,
    "end": 42.28,
    "text": "That's what Samira Abidini looked into in her research, she discovered that AI agents might not behave all their different from us humans afterall!"
  },
  {
    "start": 42.94,
    "end": 73.14,
    "text": "So let's dive IN!"
  },
  {
    "start": 78.399,
    "end": 80.66,
    "text": "And trust me, I'm not a computer scientist either."
  },
  {
    "start": 80.92,
    "end": 82.32,
    "text": "So i am learning right along with you!"
  },
  {
    "start": 82.82,
    "end": 86.699,
    "text": "Today we're going to talk about AI agents and how they arrive at decisions together."
  },
  {
    "start": 87.02,
    "end": 91.64,
    "text": "Turns out that decision finding process can be hijacked and introduce some brand new security risks."
  },
  {
    "start": 92.46,
    "end": 99.479,
    "text": "Samira Abedini master student working at CISPA and soon-to-be PhD candidate here has looked into this topic so I invited her here."
  },
  {
    "start": 99.619,
    "end": 100.979,
    "text": "Hey samirah glad to have you on board."
  },
  {
    "start": 101.199,
    "end": 101.36,
    "text": "Hi"
  },
  {
    "start": 101.42,
    "end": 103.6,
    "text": "Tobias thank you for having today"
  },
  {
    "start": 104.699,
    "end": 105.44,
    "text": "To get us started."
  },
  {
    "start": 106.38,
    "end": 110.8,
    "text": "let's... Short to talk about what an AI agent actually is and what it does."
  },
  {
    "start": 111.32,
    "end": 116.28,
    "text": "I think by now we've probably all interacted with one of these agents, but i'm sure you can give us a better idea."
  },
  {
    "start": 116.82,
    "end": 117.259,
    "text": "Yeah Sure."
  },
  {
    "start": 117.72,
    "end": 119.399,
    "text": "So let me start with an example."
  },
  {
    "start": 119.559,
    "end": 122.44,
    "text": "We all have interacted with chat gpt for example."
  },
  {
    "start": 122.559,
    "end": 122.66,
    "text": "What?"
  },
  {
    "start": 122.74,
    "end": 125.08,
    "text": "We call the chatbot or chat assistant."
  },
  {
    "start": 125.179,
    "end": 128.48,
    "text": "We just write a question And then it gives us an answer."
  },
  {
    "start": 128.68,
    "end": 134.04,
    "text": "For example if you want to go up to a vacation plan for it."
  },
  {
    "start": 134.44,
    "end": 139.62,
    "text": "You ask the charter system, and you can do this or that."
  },
  {
    "start": 140.28,
    "end": 145.6,
    "text": "but in case of AI agent we have a broader perspective."
  },
  {
    "start": 145.68,
    "end": 155.42,
    "text": "Like this AI agent can look for websites, find hotels where you maybe if you give access to some calendar or your email."
  },
  {
    "start": 155.84,
    "end": 166.74,
    "text": "it can also add the time of vacation to your calendar and it can track the progress of this search as a memory make decisions."
  },
  {
    "start": 167.04,
    "end": 169.88,
    "text": "so we have more autonomy in AI agents."
  },
  {
    "start": 169.98,
    "end": 170.32,
    "text": "I could say"
  },
  {
    "start": 170.48,
    "end": 170.88,
    "text": "Mm-hmm."
  },
  {
    "start": 171.14,
    "end": 175.74,
    "text": "So, so chat GPT is basically a very powerful agent that can do everything for you."
  },
  {
    "start": 176.04,
    "end": 180.04,
    "text": "if you saw everything in yeah"
  },
  {
    "start": 180.08,
    "end": 190.92,
    "text": "Yeah, maybe we can say the open AI agent Can't do more things where you and then they chat assistant?"
  },
  {
    "start": 191.36,
    "end": 191.46,
    "text": "Yeah"
  },
  {
    "start": 191.78,
    "end": 193.52,
    "text": "Okay And it's assistant would be something like."
  },
  {
    "start": 193.82,
    "end": 193.96,
    "text": "I mean."
  },
  {
    "start": 194.02,
    "end": 195.78,
    "text": "You mentioned uh travel example."
  },
  {
    "start": 195.9,
    "end": 199.52,
    "text": "If i go to an airline website They usually have these."
  },
  {
    "start": 200.06,
    "end": 200.92,
    "text": "Do you have any questions?"
  },
  {
    "start": 200.98,
    "end": 205.64,
    "text": "And then it's also just like an AI agent, but basically customer service thing."
  },
  {
    "start": 205.68,
    "end": 215.68,
    "text": "Yeah exactly with AI agents we mean... It can make decisions, how goals and track the progress of some tasks."
  },
  {
    "start": 216.2,
    "end": 221.76,
    "text": "Okay in your paper focus on having multiple AI agents working together."
  },
  {
    "start": 221.9,
    "end": 224.2,
    "text": "so what would be upside if something like that?"
  },
  {
    "start": 224.56,
    "end": 228.16,
    "text": "Yes, I think...I look from human perspective here."
  },
  {
    "start": 229.0,
    "end": 237.8,
    "text": "The idea is similar to why we have group of people working on a task and the same idea applies to AI agents for example."
  },
  {
    "start": 239.08,
    "end": 251.04,
    "text": "One advantage as we can divide tasks into some small parts And then finish them in parallel so that they hold big task finished soon."
  },
  {
    "start": 251.18,
    "end": 251.86,
    "text": "We are faster."
  },
  {
    "start": 252.52,
    "end": 258.42,
    "text": "Or maybe like the humans, we have different perspectives when working in groups together."
  },
  {
    "start": 259.18,
    "end": 270.18,
    "text": "If you have heterogeneous agents in this multi-agent setup each agent can come up from a different perspective then outcome will be much better for us."
  },
  {
    "start": 270.98,
    "end": 275.14,
    "text": "or also make each agent more specialized"
  },
  {
    "start": 275.32,
    "end": 275.38,
    "text": "e.g.,"
  },
  {
    "start": 276.16,
    "end": 277.1,
    "text": "encoding agents."
  },
  {
    "start": 277.98,
    "end": 281.8,
    "text": "We have a specialized agent for writing the code, want to review the court."
  },
  {
    "start": 282.32,
    "end": 286.2,
    "text": "Wanted to execute it or went through search for websites and stuff."
  },
  {
    "start": 286.94,
    "end": 288.46,
    "text": "making them specialized."
  },
  {
    "start": 289.58,
    "end": 291.6,
    "text": "help us more I think."
  },
  {
    "start": 292.38,
    "end": 296.32,
    "text": "so you give each agent basically more narrow field of expertise."
  },
  {
    "start": 296.56,
    "end": 301.14,
    "text": "So they basically have different strength as he would have in human team."
  },
  {
    "start": 301.48,
    "end": 302.08,
    "text": "Yeah exactly"
  },
  {
    "start": 302.38,
    "end": 303.08,
    "text": "mm-hmm okay."
  },
  {
    "start": 303.94,
    "end": 313.76,
    "text": "But your paper also introduces the idea that if we have these several AI agents working together, then we might actually create new security risks than we haven't really thought about yet."
  },
  {
    "start": 313.88,
    "end": 316.7,
    "text": "So how did you first become interested in this problem?"
  },
  {
    "start": 317.5,
    "end": 339.34,
    "text": "Yeah so I first came across with an idea of different LLM stakeholders negotiating each other on one paper by Sahara Abdel Nabi at all and like it was about all of them negotiating with each other and they were analyzing if some competition happens in this negotiation."
  },
  {
    "start": 340.1,
    "end": 345.78,
    "text": "And then suddenly I thought, oh these agents are not just interacting with humans always?"
  },
  {
    "start": 346.02,
    "end": 360.68,
    "text": "They may be at some point they will interact with eachother... ...and what happens if they interact with one another on a communication channel that we built to make the system if more efficient or get some advantage out of it."
  },
  {
    "start": 360.82,
    "end": 367.52,
    "text": "It can also turn to be the same channel, spread vulnerability and misinformation."
  },
  {
    "start": 368.36,
    "end": 371.66,
    "text": "So my interest started from that point."
  },
  {
    "start": 372.58,
    "end": 379.68,
    "text": "The idea is the more AI we got there... ...the more AI takes over human jobs."
  },
  {
    "start": 380.28,
    "end": 382.16,
    "text": "That's not called jobs but tasks."
  },
  {
    "start": 383.76,
    "end": 385.56,
    "text": "They will interact with each other instead."
  },
  {
    "start": 386.52,
    "end": 388.4,
    "text": "Humans versus humans with AI."
  },
  {
    "start": 388.88,
    "end": 401.14,
    "text": "I can say it's going to be more hard for us To see what's happening behind like if some agents are communicating With each other to do a task at some point."
  },
  {
    "start": 401.52,
    "end": 402.64,
    "text": "Do we have enough time?"
  },
  {
    "start": 402.7,
    "end": 407.44,
    "text": "To read all the traces to see What my agent told your agent and the other agent?"
  },
  {
    "start": 408.1,
    "end": 410.68,
    "text": "so yeah, this is The problem."
  },
  {
    "start": 411.72,
    "end": 414.4,
    "text": "okay So tell us what was the problem exactly?"
  },
  {
    "start": 414.44,
    "end": 419.3,
    "text": "What did you find out, why could make a multi-agent system set up dangerous?"
  },
  {
    "start": 419.72,
    "end": 420.0,
    "text": "let's say."
  },
  {
    "start": 420.62,
    "end": 430.1,
    "text": "Yeah so the main point of having this multi agent setup is creating communication between different agents and get advantage over it in that way."
  },
  {
    "start": 430.8,
    "end": 437.36,
    "text": "but same communication channel can use to spread some vulnerabilities, as I mentioned before."
  },
  {
    "start": 438.22,
    "end": 440.3,
    "text": "Or some misinformation for example."
  },
  {
    "start": 440.64,
    "end": 445.02,
    "text": "and the problem is that bigger system becomes just a small."
  },
  {
    "start": 445.24,
    "end": 448.6,
    "text": "kind of vulnerabilities will go through whole network."
  },
  {
    "start": 449.12,
    "end": 451.08,
    "text": "suddenly it become bigger."
  },
  {
    "start": 453.0,
    "end": 454.76,
    "text": "this is main problems."
  },
  {
    "start": 455.44,
    "end": 464.82,
    "text": "if one agent starts to pollute its neighbor then the other neighbor start to pollutes own neighbors and this will go on in the network."
  },
  {
    "start": 465.36,
    "end": 466.92,
    "text": "This is the main problem, my"
  },
  {
    "start": 466.98,
    "end": 467.3,
    "text": "opinion.\"."
  },
  {
    "start": 467.94,
    "end": 472.48,
    "text": "Okay so maybe let's take a step back like please kind of describe your trial setup."
  },
  {
    "start": 472.56,
    "end": 475.7,
    "text": "but you have different agents discussing a problem."
  },
  {
    "start": 476.52,
    "end": 477.84,
    "text": "what does it look like in practice?"
  },
  {
    "start": 477.94,
    "end": 478.82,
    "text": "Like what are they discussing."
  },
  {
    "start": 479.04,
    "end": 479.9,
    "text": "how can we imagine"
  },
  {
    "start": 484.8,
    "end": 492.48,
    "text": "So in the paper setup we actually test different agents, different topology of networks and different tasks."
  },
  {
    "start": 493.02,
    "end": 501.54,
    "text": "But like tasks are from two callings to selecting that best tool to call some common sense question answered."
  },
  {
    "start": 501.94,
    "end": 505.14,
    "text": "but let me give you an example with a common sense questions answer."
  },
  {
    "start": 506.1,
    "end": 509.3,
    "text": "So imagine we have six agents at the beginning."
  },
  {
    "start": 509.4,
    "end": 512.86,
    "text": "We give a question like, what is someone trying for?"
  },
  {
    "start": 512.919,
    "end": 516.62,
    "text": "Someone who's jogging What's trying to reach?"
  },
  {
    "start": 517.02,
    "end": 518.6,
    "text": "or was it the goal?"
  },
  {
    "start": 518.84,
    "end": 538.96,
    "text": "yeah where some of us join and then each of the agents given an initial answer without any interaction with other agents And after they have an answer, a reasoning why we chose this answer and then how confident you are in the answer."
  },
  {
    "start": 539.82,
    "end": 544.36,
    "text": "After that they spread these information to their neighbors."
  },
  {
    "start": 544.64,
    "end": 548.46,
    "text": "They give them reason for their neighbors And save it from memory."
  },
  {
    "start": 549.66,
    "end": 552.82,
    "text": "The next round of teleportation starts."
  },
  {
    "start": 553.28,
    "end": 558.12,
    "text": "The agents get answers And also they have their own answer."
  },
  {
    "start": 558.54,
    "end": 561.0,
    "text": "So, they had to see what's the situation?"
  },
  {
    "start": 561.4,
    "end": 562.42,
    "text": "This was my opinion."
  },
  {
    "start": 562.5,
    "end": 570.06,
    "text": "and now this is another agent saying that... ...and then update their answers based on neighbors' memories of theirs."
  },
  {
    "start": 570.16,
    "end": 576.3,
    "text": "They come up with an answer a reasoning again spread through the neighbours'."
  },
  {
    "start": 576.72,
    "end": 582.14,
    "text": "And so several times we've got communication happening."
  },
  {
    "start": 582.8,
    "end": 586.12,
    "text": "at that point one of agents might Have a wrong answer."
  },
  {
    "start": 587.24,
    "end": 588.78,
    "text": "Okay, and what happens then?"
  },
  {
    "start": 589.54,
    "end": 595.66,
    "text": "So let's say in your example like the main goal common sense would be okay."
  },
  {
    "start": 595.82,
    "end": 598.82,
    "text": "someone is trying to get fit so I'm just going to get out of this way."
  },
  {
    "start": 598.96,
    "end": 599.74,
    "text": "maybe they are trying too."
  },
  {
    "start": 600.52,
    "end": 603.44,
    "text": "yeah we want to get faster fitness as the gold basically."
  },
  {
    "start": 603.56,
    "end": 608.14,
    "text": "so Let's see one agent says The Gold is two...I don't know"
  },
  {
    "start": 608.88,
    "end": 610.0,
    "text": "To gain some pain"
  },
  {
    "start": 611.28,
    "end": 612.8,
    "text": "to get larger feet, whatever."
  },
  {
    "start": 613.94,
    "end": 615.14,
    "text": "So what happens then?"
  },
  {
    "start": 616.16,
    "end": 621.04,
    "text": "Yeah this is the main problem that we have in The same task."
  },
  {
    "start": 621.16,
    "end": 631.14,
    "text": "We had options like okay This one person wants to become fit or healthier But if he also had options at the person tries to gain some feet pain."
  },
  {
    "start": 631.64,
    "end": 632.08,
    "text": "Okay"
  },
  {
    "start": 632.48,
    "end": 637.12,
    "text": "We asked the attacker agent in the system prompt to defend this answer."
  },
  {
    "start": 637.76,
    "end": 649.68,
    "text": "And then, okay at first round of communication maybe it can change other agents' beliefs I could say but they make them less confident than their own answers."
  },
  {
    "start": 650.4,
    "end": 661.54,
    "text": "and when these agents gives a wrong answer for each other and becomes less confident In next round The other agent is also source of misinformation."
  },
  {
    "start": 662.54,
    "end": 667.86,
    "text": "not only the attacker is spreading this, or that person's trying to have pain."
  },
  {
    "start": 668.32,
    "end": 671.34,
    "text": "The other agent has also somehow tried to say it."
  },
  {
    "start": 671.86,
    "end": 679.6,
    "text": "and then when we go through the deliberation both of them become more confident And they affect their own neighbors."
  },
  {
    "start": 682.16,
    "end": 688.38,
    "text": "We call cascade effect at end which just a small attack starts and becomes bigger"
  },
  {
    "start": 692.5,
    "end": 697.8,
    "text": "one agent having a clearly wrong opinion and propagating it through the network."
  },
  {
    "start": 698.04,
    "end": 703.22,
    "text": "Yes, exactly that misinformation is trying to propagate in the network"
  },
  {
    "start": 703.4,
    "end": 703.6,
    "text": "okay?"
  },
  {
    "start": 703.7,
    "end": 705.22,
    "text": "And then you only need One Agent."
  },
  {
    "start": 706.08,
    "end": 706.88,
    "text": "That Is Clearly Wrong."
  },
  {
    "start": 706.96,
    "end": 714.42,
    "text": "and Then They Can Influence A Whole Network and In The End The Whole Network Might Agree On Or Clearly Wrong Answer"
  },
  {
    "start": 714.72,
    "end": 718.76,
    "text": "yes if the agent is stubborn enough To Do This."
  },
  {
    "start": 719.44,
    "end": 723.46,
    "text": "Okay Stubborn Like That's a good word."
  },
  {
    "start": 723.78,
    "end": 728.06,
    "text": "What does it mean when an agent is stubborn, which is kind of human quality?"
  },
  {
    "start": 728.7,
    "end": 731.34,
    "text": "How would you describe this in terms of AI agents?"
  },
  {
    "start": 731.4,
    "end": 740.32,
    "text": "I want to say we are not associating human behavior exactly with the agents but this easier way for us to describe and feature that agents."
  },
  {
    "start": 740.4,
    "end": 742.8,
    "text": "how what they mean by stubbornness?"
  },
  {
    "start": 742.94,
    "end": 751.12,
    "text": "as you have initial belief in one answer and whatever the communication is happening, you're not willing to change your answer."
  },
  {
    "start": 751.84,
    "end": 754.8,
    "text": "And next round we checked answer it's same again."
  },
  {
    "start": 754.92,
    "end": 764.96,
    "text": "Next round We check that sir It's the same Again if agent like this because stubborn yeah somehow like some humans Yeah"
  },
  {
    "start": 765.0,
    "end": 769.16,
    "text": "well I can tell You i have met Humans before That were exactly Like that."
  },
  {
    "start": 769.24,
    "end": 772.64,
    "text": "so So that makes sense."
  },
  {
    "start": 774.2,
    "end": 776.88,
    "text": "So just to sum this up, we have like one stubborn attacker."
  },
  {
    "start": 777.16,
    "end": 778.84,
    "text": "Um which is a?"
  },
  {
    "start": 779.56,
    "end": 786.24,
    "text": "basically let's call it a rogue agent that has wrong opinion and Let's say their goal was to misinform."
  },
  {
    "start": 787.02,
    "end": 787.82,
    "text": "And they only have two."
  },
  {
    "start": 788.8,
    "end": 795.48,
    "text": "Basically turn One Agent next of them and from then It can cascade through."
  },
  {
    "start": 795.52,
    "end": 800.6,
    "text": "the whole network will not have a correct answer in the end."
  },
  {
    "start": 800.78,
    "end": 800.94,
    "text": "Yes,"
  },
  {
    "start": 801.12,
    "end": 810.52,
    "text": "exactly so that this agent if it's like has these features stubbornness and It can influence their neighbors."
  },
  {
    "start": 811.08,
    "end": 817.6,
    "text": "then the neighbors are affected than the neighbors will Influenced others as well And at the end we would have a cascade effect."
  },
  {
    "start": 817.7,
    "end": 823.76,
    "text": "and then The outcome of our multi-agent system is steered towards the attackers."
  },
  {
    "start": 824.32,
    "end": 824.68,
    "text": "opinion"
  },
  {
    "start": 825.06,
    "end": 825.22,
    "text": "mm-hmm."
  },
  {
    "start": 825.6,
    "end": 830.62,
    "text": "okay Okay, so in this field we always call these things attacks."
  },
  {
    "start": 830.8,
    "end": 832.28,
    "text": "I mean which they of course are."
  },
  {
    "start": 832.58,
    "end": 845.36,
    "text": "but i think because we talk about attacks and vulnerabilities the question is it's not really exploiting a technical vulnerability per se It simply doing what its supposed to do right?"
  },
  {
    "start": 845.44,
    "end": 851.9,
    "text": "They're discussing like one agent argument just stronger even if their wrong."
  },
  {
    "start": 853.38,
    "end": 858.94,
    "text": "Why is that important from a security perspective if there's no exploit but more like system design thing?"
  },
  {
    "start": 859.36,
    "end": 860.04,
    "text": "Yeah, exactly."
  },
  {
    "start": 860.08,
    "end": 861.7,
    "text": "I think this is very nice question."
  },
  {
    "start": 861.8,
    "end": 870.54,
    "text": "actually as you mentioned when we talk about attacks it almost always about exploiting something box, I don't know."
  },
  {
    "start": 871.1,
    "end": 882.72,
    "text": "So injecting some malicious instructions or jailbreaking are trying to make an agent say break the safeguards."
  },
  {
    "start": 883.34,
    "end": 889.82,
    "text": "but here in our case The system is designed this way and attacker also following their rules."
  },
  {
    "start": 889.9,
    "end": 896.94,
    "text": "it's not breaking anything for example changing topology or change the deliberation protocol Nothing."
  },
  {
    "start": 897.58,
    "end": 902.48,
    "text": "So the design is safe, if you can say it."
  },
  {
    "start": 902.86,
    "end": 912.06,
    "text": "but the same communication channel can also be used strategically by some agents through the same language to steer that outcome."
  },
  {
    "start": 912.56,
    "end": 923.06,
    "text": "This is exactly why its important from security perspective because It's not enough to say okay all the nodes are safe so this system is safe."
  },
  {
    "start": 924.52,
    "end": 932.5,
    "text": "We should also pay attention to the communication happening, and this is really important from a security perspective in multi-agent systems."
  },
  {
    "start": 933.76,
    "end": 942.44,
    "text": "And if it's happening through these cascading process just like adding more AI agents... ...is not going to solve the problem initially right?"
  },
  {
    "start": 942.52,
    "end": 944.98,
    "text": "So you have to think of something else to defend against."
  },
  {
    "start": 945.42,
    "end": 948.28,
    "text": "Yeah exactly adding more agents."
  },
  {
    "start": 950.26,
    "end": 960.54,
    "text": "The insights we get from our research is that topology is important, where you put the agents are important or how they're communicating."
  },
  {
    "start": 961.06,
    "end": 965.06,
    "text": "all of these are important to have a safe multi-agent setup."
  },
  {
    "start": 965.5,
    "end": 967.5,
    "text": "so it's not just enough okay?"
  },
  {
    "start": 967.78,
    "end": 970.28,
    "text": "We want to have an intelligent system."
  },
  {
    "start": 970.4,
    "end": 972.26,
    "text": "let us put lots of agents together."
  },
  {
    "start": 972.56,
    "end": 975.4,
    "text": "maybe one of those agents are malicious for example."
  },
  {
    "start": 976.02,
    "end": 977.58,
    "text": "Okay, you talk about the topology."
  },
  {
    "start": 977.7,
    "end": 978.78,
    "text": "Like what exactly does that mean?"
  },
  {
    "start": 978.82,
    "end": 983.82,
    "text": "Does it basically matter where like how one agent talks to another?"
  },
  {
    "start": 984.1,
    "end": 986.06,
    "text": "or maybe can describe something?"
  },
  {
    "start": 986.76,
    "end": 992.3,
    "text": "so by topology I mean the design of network we have."
  },
  {
    "start": 992.9,
    "end": 999.04,
    "text": "For example, in the paper we discussed a fully connected networks which all the agents are communicating with."
  },
  {
    "start": 999.26,
    "end": 1000.3,
    "text": "All day other agents."
  },
  {
    "start": 1000.78,
    "end": 1016.22,
    "text": "or We have star topology Which is also very famous and multi agent setup that we Have an Agent In The Middle As A Hub And Then We Have The Leaves Which Just Talk To The Hub Agent Not Any Agents."
  },
  {
    "start": 1016.28,
    "end": 1018.9,
    "text": "So We Call These Designs Topology."
  },
  {
    "start": 1020.0,
    "end": 1024.359,
    "text": "And Yeah for Example The Star Topology is vulnerable."
  },
  {
    "start": 1024.46,
    "end": 1027.3,
    "text": "If the hub is malicious, you"
  },
  {
    "start": 1028.0,
    "end": 1037.48,
    "text": "can not as susceptible if there's like someone on the outer branches right because the hub can more easily distinguish between what is right and"
  },
  {
    "start": 1037.52,
    "end": 1038.319,
    "text": "wrong exactly?"
  },
  {
    "start": 1038.46,
    "end": 1041.339,
    "text": "Exactly so that this Is a very good point!"
  },
  {
    "start": 1041.68,
    "end": 1044.94,
    "text": "If the attacker isn't alive then The system is More."
  },
  {
    "start": 1045.96,
    "end": 1051.08,
    "text": "It's Not That Vulnerable to the attack but when the Hub is hijacked it's more vulnerable."
  },
  {
    "start": 1051.66,
    "end": 1054.02,
    "text": "okay Okay, good."
  },
  {
    "start": 1054.32,
    "end": 1058.1,
    "text": "So that was quite technical but I think also quite accessible."
  },
  {
    "start": 1058.2,
    "end": 1061.22,
    "text": "so... But you're also to understand these interactions."
  },
  {
    "start": 1061.28,
    "end": 1065.6,
    "text": "borrow the model from social sciences which i thought it's interesting."
  },
  {
    "start": 1066.42,
    "end": 1070.14,
    "text": "It is called Friedkin Johnson Model of Opinion Formation."
  },
  {
    "start": 1071.28,
    "end": 1073.04,
    "text": "Not very easy to remember."
  },
  {
    "start": 1075.98,
    "end": 1078.02,
    "text": "How did get the idea to do that."
  },
  {
    "start": 1078.06,
    "end": 1079.14,
    "text": "Why did you choose a model?"
  },
  {
    "start": 1079.24,
    "end": 1082.54,
    "text": "That was originally developed, so study how humans influence each other"
  },
  {
    "start": 1082.86,
    "end": 1084.0,
    "text": "as we discussed with them."
  },
  {
    "start": 1084.16,
    "end": 1094.72,
    "text": "means of communication is the same language that we humans are using and also The thing which is Going through it's the information."
  },
  {
    "start": 1094.8,
    "end": 1100.9,
    "text": "We're passing their in formation two other people an agents or in multi agent setup As well."
  },
  {
    "start": 1101.24,
    "end": 1103.4,
    "text": "there they're spreading the information already."
  },
  {
    "start": 1103.44,
    "end": 1106.76,
    "text": "They're passing the information In the network"
  },
  {
    "start": 1107.4,
    "end": 1108.68,
    "text": "But I think that's an important point."
  },
  {
    "start": 1108.8,
    "end": 1111.36,
    "text": "So they're doing it in like plain human language?"
  },
  {
    "start": 1111.62,
    "end": 1111.78,
    "text": "Yes,"
  },
  {
    "start": 1112.14,
    "end": 1112.96,
    "text": "okay yeah."
  },
  {
    "start": 1113.24,
    "end": 1125.34,
    "text": "so we thought Okay if the goal is passing information and the language Is same can you see If The Same Concepts Can Be Adopted To These Agents?"
  },
  {
    "start": 1125.8,
    "end": 1133.2,
    "text": "And Then We Just Started to Explore and We Found That Fritkin Janssen Model Which We Also Call It Fj-Model."
  },
  {
    "start": 1134.22,
    "end": 1134.8,
    "text": "Yeah,"
  },
  {
    "start": 1138.4,
    "end": 1140.24,
    "text": "just to model the interaction between agents."
  },
  {
    "start": 1140.48,
    "end": 1146.36,
    "text": "Okay so but as we said before AI agents are not very similar to humans."
  },
  {
    "start": 1146.46,
    "end": 1151.96,
    "text": "But what does the behavior of AI agents have in common with how humans arrive at their opinions and groups?"
  },
  {
    "start": 1152.22,
    "end": 1155.84,
    "text": "So let me get first go through this FGA model."
  },
  {
    "start": 1156.64,
    "end": 1160.08,
    "text": "which different parts it has then be easier for us to understand."
  },
  {
    "start": 1160.24,
    "end": 1163.36,
    "text": "So FGA Model For Humans It Has Three Parts."
  },
  {
    "start": 1163.9,
    "end": 1173.28,
    "text": "The First part is about modeling the stubbornness And which means how a person sticks to its own initial opinion."
  },
  {
    "start": 1173.76,
    "end": 1188.6,
    "text": "The next part, the second part is about how it resists to change their opinions around an opinion and at that end we have... How they get influenced by other people for example."
  },
  {
    "start": 1188.7,
    "end": 1190.14,
    "text": "So these are three parts."
  },
  {
    "start": 1190.54,
    "end": 1201.42,
    "text": "The stubbornness, resistance they have and influence some kind of same concepts in the multi-agent set up with agents."
  },
  {
    "start": 1202.36,
    "end": 1207.42,
    "text": "The idea is we want to see if an agent changes initial belief or not, okay?"
  },
  {
    "start": 1208.16,
    "end": 1215.6,
    "text": "If it gets influenced by other agents or not... ...if it wants to change its last round leaf for example or not."
  },
  {
    "start": 1216.3,
    "end": 1218.98,
    "text": "So this what's common?"
  },
  {
    "start": 1219.7,
    "end": 1221.4,
    "text": "these two concepts?"
  },
  {
    "start": 1221.78,
    "end": 1230.22,
    "text": "Okay so basically were talking about how stubborn are they And how trusting are they when someone else tries to tell them something?"
  },
  {
    "start": 1230.72,
    "end": 1235.12,
    "text": "Maybe we can say influence, like how they're influenced by the others."
  },
  {
    "start": 1235.24,
    "end": 1236.7,
    "text": "If there's somebody saying something"
  },
  {
    "start": 1237.42,
    "end": 1241.98,
    "text": "and then you found out that this works also quite well for AI agents right."
  },
  {
    "start": 1242.3,
    "end": 1254.08,
    "text": "so You couldn't say an agent is very stubborn it sticks to its own opinion and maybe another one Is very let's say easily influenced By The other model."
  },
  {
    "start": 1255.66,
    "end": 1259.56,
    "text": "Okay, and what did you do with these findings?"
  },
  {
    "start": 1259.62,
    "end": 1266.44,
    "text": "How did you apply them to your multi-AI agents?"
  },
  {
    "start": 1269.46,
    "end": 1270.88,
    "text": "To your multi agent setups."
  },
  {
    "start": 1271.02,
    "end": 1278.62,
    "text": "So yeah so first when we fitted this model on our data we got from the deliberation in multiagent setup."
  },
  {
    "start": 1278.76,
    "end": 1280.0,
    "text": "We found that okay."
  },
  {
    "start": 1280.32,
    "end": 1288.6,
    "text": "When we have a particularly stubborn agent in the network then we fit the parameter which is associated with stubbornness."
  },
  {
    "start": 1288.7,
    "end": 1295.96,
    "text": "It's really high like when they agent and it trying not to change his opinion, And this was an insight."
  },
  {
    "start": 1296.56,
    "end": 1297.1,
    "text": "we got that."
  },
  {
    "start": 1297.2,
    "end": 1303.2,
    "text": "okay yeah This model maybe can To some degree explain how is this communication happening?"
  },
  {
    "start": 1304.74,
    "end": 1312.14,
    "text": "Mm-hmm best insights we got from this was can we also come up with some mitigation approaches based on the insights?"
  },
  {
    "start": 1312.2,
    "end": 1312.5,
    "text": "We get it."
  },
  {
    "start": 1312.8,
    "end": 1324.42,
    "text": "From the model and This is what's the point that visa focused in influence part said, okay if they other agents isn't malicious and its influencing our agent."
  },
  {
    "start": 1324.92,
    "end": 1328.16,
    "text": "What if we decrease this influence."
  },
  {
    "start": 1328.48,
    "end": 1341.22,
    "text": "And then the agent can detect if there's a wrong agent and decreased its influence, than also inspired by social sciences called distrust actually."
  },
  {
    "start": 1342.08,
    "end": 1355.84,
    "text": "So you basically decided that you have to focus on part of your control which is how easily influences my AI agent because I don't know when i put it out there with other agents what they are Telling it"
  },
  {
    "start": 1355.92,
    "end": 1357.78,
    "text": "basically yeah exactly"
  },
  {
    "start": 1358.16,
    "end": 1359.2,
    "text": "okay interesting so."
  },
  {
    "start": 1360.38,
    "end": 1361.58,
    "text": "I mean we've shortly talked about."
  },
  {
    "start": 1361.88,
    "end": 1367.14,
    "text": "its not enough to put in more agent into a multi-agent model too."
  },
  {
    "start": 1367.18,
    "end": 1375.64,
    "text": "defend against these attacks, but this is actually something that can defend against his attacks and if you kind of harden them against influence by rogue agents."
  },
  {
    "start": 1376.14,
    "end": 1376.8,
    "text": "Yes, exactly."
  },
  {
    "start": 1377.56,
    "end": 1391.34,
    "text": "If this trust metric is calibrated correctly then the agent can find out which agents I could get influenced by and which other agents i shouldn't be influenced by."
  },
  {
    "start": 1391.6,
    "end": 1393.42,
    "text": "Okay but it sounds like there's a trade-off right?"
  },
  {
    "start": 1394.06,
    "end": 1396.32,
    "text": "Let say you're not influenced very easily."
  },
  {
    "start": 1397.38,
    "end": 1403.26,
    "text": "It also gets to that point where you are not influenced even if YOU are wrong."
  },
  {
    "start": 1404.1,
    "end": 1406.36,
    "text": "yes This is a very good point."
  },
  {
    "start": 1406.44,
    "end": 1422.24,
    "text": "There's the trade-off there, The main point of having multiagent systems is collaboration and communication And if we just decrease this influence it means that We are not collaborating anymore Which doesn't make sense for a multi agent system."
  },
  {
    "start": 1423.14,
    "end": 1426.4,
    "text": "So its really important to find this trade off."
  },
  {
    "start": 1426.98,
    "end": 1431.38,
    "text": "when to trust to who and when not to trust whom."
  },
  {
    "start": 1431.5,
    "end": 1435.78,
    "text": "And, uh... Not make it strict that okay I don't hear anything."
  },
  {
    "start": 1436.14,
    "end": 1439.4,
    "text": "so yeah then the math agent system doesn't"
  },
  {
    "start": 1439.44,
    "end": 1439.86,
    "text": "makes sense."
  },
  {
    "start": 1440.1,
    "end": 1442.16,
    "text": "So you need find a sweet spot basically?"
  },
  {
    "start": 1442.44,
    "end": 1448.86,
    "text": "Exactly this is the case with the trust defense we have for some time."
  },
  {
    "start": 1449.0,
    "end": 1450.22,
    "text": "We calibrate trust."
  },
  {
    "start": 1450.82,
    "end": 1456.44,
    "text": "after they let agents start communicate with discalibrated trust"
  },
  {
    "start": 1456.8,
    "end": 1459.0,
    "text": "Okay And like, what is your approach to that?"
  },
  {
    "start": 1459.08,
    "end": 1459.46,
    "text": "How do you?"
  },
  {
    "start": 1460.28,
    "end": 1462.02,
    "text": "would she do it in the real world."
  },
  {
    "start": 1462.54,
    "end": 1471.04,
    "text": "Yeah for example In Real World I can say maybe we have some ground truth tasks For example."
  },
  {
    "start": 1471.44,
    "end": 1476.26,
    "text": "then We Can Put our agent and a multi-agent setup with these Ground Truths in mind."
  },
  {
    "start": 1476.62,
    "end": 1480.42,
    "text": "yeah and the agent interacts with other agents."
  },
  {
    "start": 1480.78,
    "end": 1492.66,
    "text": "And whenever this, uh... The answer of Other Agents is not based on this ground truth that we have then they know if it repeats something wrong to an agent I shouldn't trust them."
  },
  {
    "start": 1492.8,
    "end": 1500.22,
    "text": "but If the other agents are collaborating in a positive way It can keep high-trusts for them."
  },
  {
    "start": 1500.58,
    "end": 1502.32,
    "text": "This way you can collaborate your trust."
  },
  {
    "start": 1502.52,
    "end": 1507.1,
    "text": "Then We start doing for that real tasks."
  },
  {
    "start": 1507.84,
    "end": 1517.62,
    "text": "Yeah, it makes sense by thinking this further because we're always doing the cat and mouse game like an attacker could also try to circumvent this trust?"
  },
  {
    "start": 1518.02,
    "end": 1527.0,
    "text": "Yes exactly so this is a case that we also tested in the paper That's an adaptive attacker can actually exploit this really well."
  },
  {
    "start": 1527.46,
    "end": 1537.68,
    "text": "Because if they know that They are being tested at these ground rules tests tasks, they can easily gain trust and then in the main deliberation."
  },
  {
    "start": 1537.76,
    "end": 1540.22,
    "text": "They can game that system."
  },
  {
    "start": 1540.36,
    "end": 1563.62,
    "text": "so this is why we also went further and improved our trust defense to have some sparse updates between In case an adaptive attacker did this, in the main deliberation we can find out which agent was this adaptive attacker and then decrease their trust to that."
  },
  {
    "start": 1565.08,
    "end": 1565.12,
    "text": "So"
  },
  {
    "start": 1565.2,
    "end": 1565.66,
    "text": "you thought ahead?"
  },
  {
    "start": 1566.38,
    "end": 1566.72,
    "text": "That's good!"
  },
  {
    "start": 1568.44,
    "end": 1580.04,
    "text": "Okay but before we see your approach into real world maybe let's just talk about like multi-AI agents in the real world where do you expect these kind of multi-agent systems become relevant?"
  },
  {
    "start": 1580.2,
    "end": 1581.1,
    "text": "or in the Real World?"
  },
  {
    "start": 1581.14,
    "end": 1584.42,
    "text": "because right now It still sounds a little sci-fi."
  },
  {
    "start": 1585.16,
    "end": 1587.06,
    "text": "Although it probably is not anymore, so"
  },
  {
    "start": 1587.7,
    "end": 1598.84,
    "text": "yeah actually I feel its not science fiction any more because there are some parts in real world that these agents are implemented."
  },
  {
    "start": 1600.14,
    "end": 1611.96,
    "text": "For example, in coding agents we have some orchestrator agent which has sub-agents and then they are more specialized."
  },
  {
    "start": 1612.14,
    "end": 1617.8,
    "text": "They go do the task And then the orchestrator sees or checks reviews of tasks."
  },
  {
    "start": 1618.94,
    "end": 1630.66,
    "text": "I think at this point We have multi-agent's In real world but also near future size of network will become bigger and bigger."
  },
  {
    "start": 1631.6,
    "end": 1647.58,
    "text": "For example, I can also give them multiple examples for instance which was social media for agents who were using OpenCloud And we saw in that incident there are a lot."
  },
  {
    "start": 1647.76,
    "end": 1651.4,
    "text": "stop happening if these agents just start to communicate with each other."
  },
  {
    "start": 1651.92,
    "end": 1661.7,
    "text": "it's An example of agent network, a huge agent network which I think in the near future will even be bigger and bigger."
  },
  {
    "start": 1662.28,
    "end": 1675.66,
    "text": "Maybe some people who have some personal agents when they write you an email maybe your agent would respond to me Yeah, and also organizations might start to use these agents."
  },
  {
    "start": 1675.74,
    "end": 1677.76,
    "text": "For example IT department has an agent."
  },
  {
    "start": 1679.16,
    "end": 1679.46,
    "text": "no more."
  },
  {
    "start": 1679.56,
    "end": 1686.24,
    "text": "humans are not answering all the tickets for example in HR department can also have an agent for themselves."
  },
  {
    "start": 1686.76,
    "end": 1691.54,
    "text": "And yeah I think we will have some real-world examples soon."
  },
  {
    "start": 1691.94,
    "end": 1697.88,
    "text": "And i think there risks like well definitely grow Like In your example with the coding agents."
  },
  {
    "start": 1698.38,
    "end": 1701.66,
    "text": "Let's say all the agents are still under the hood of one company."
  },
  {
    "start": 1702.48,
    "end": 1709.84,
    "text": "But once different stakeholders start communicating via AI agents, then I think kind of cats out at a back"
  },
  {
    "start": 1710.32,
    "end": 1710.74,
    "text": "exactly."
  },
  {
    "start": 1711.7,
    "end": 1733.9,
    "text": "so if i also give an example about booking system or booking agent imagine we have different hotels or different flight companies and stuff And your agent wants to book some hotels for you're benefit is not the same as hotels where they want to sell higher and you find the best hotel at a cheap price."
  },
  {
    "start": 1733.98,
    "end": 1741.64,
    "text": "Then their agent might, if there are agents... able to influence your agent, buy the highest price."
  },
  {
    "start": 1741.84,
    "end": 1745.34,
    "text": "This is a problem because we have different stakeholders with different benefits."
  },
  {
    "start": 1745.84,
    "end": 1752.74,
    "text": "also vice versa like Your agent can influence the hotel's agent maybe To find cheaper room."
  },
  {
    "start": 1753.04,
    "end": 1759.38,
    "text": "Yeah I think if you have different stake holders this will become even much more important."
  },
  {
    "start": 1759.88,
    "end": 1762.78,
    "text": "Okay Where do you think your research would go from here?"
  },
  {
    "start": 1762.92,
    "end": 1765.84,
    "text": "Like Will You keep working on that or will you change the scope?"
  },
  {
    "start": 1766.06,
    "end": 1786.52,
    "text": "yeah I will definitely continue to work on multi-agent system because i think it's important to understand how this communication is happening and if these networks grow up further in the near future, It would really be important for us to understand beforehand what was going on."
  },
  {
    "start": 1787.1,
    "end": 1793.6,
    "text": "And not have even more or much bigger incidents."
  },
  {
    "start": 1795.8,
    "end": 1802.34,
    "text": "If you could give developers of multi-agent AI systems one security recommendation right now based on your research, what would it be?"
  },
  {
    "start": 1802.62,
    "end": 1803.36,
    "text": "Maybe someone's listening."
  },
  {
    "start": 1809.08,
    "end": 1820.6,
    "text": "Okay maybe I can say topology design is important keeping each of the nodes or agents safe as important but also The communication is important."
  },
  {
    "start": 1820.76,
    "end": 1825.7,
    "text": "It's not enough to just say, okay all the agents are from this company they're safe and stuff."
  },
  {
    "start": 1825.88,
    "end": 1831.66,
    "text": "no we should also pay attention to the communication We have or which part of the network?"
  },
  {
    "start": 1831.76,
    "end": 1836.26,
    "text": "We put a really powerful agent for example as I mentioned first star."
  },
  {
    "start": 1836.4,
    "end": 1841.06,
    "text": "if you put a Powerful star a powerful agent in the hub of a start it can't."
  },
  {
    "start": 1841.34,
    "end": 1847.88,
    "text": "it Can really hijack the whole Network because the other agents don't have any other information resource."
  },
  {
    "start": 1848.74,
    "end": 1851.3,
    "text": "So it's really important to pay attention."
  },
  {
    "start": 1851.42,
    "end": 1860.82,
    "text": "To what design choice we have for the network and also making this communication safe as well, not just notes."
  },
  {
    "start": 1861.76,
    "end": 1876.3,
    "text": "I mean if you're thinking of this there could be like even a business case in someone hosting trusted agent that is serving as mediator but they would juicy target, let's say."
  },
  {
    "start": 1877.04,
    "end": 1877.48,
    "text": "Exactly!"
  },
  {
    "start": 1878.36,
    "end": 1893.96,
    "text": "So in our research we found that StarHob attacker is the most powerful attacker and I can say more of the coding multi-agent systems that i worked with were star topology."
  },
  {
    "start": 1894.86,
    "end": 1897.64,
    "text": "so maybe you should think about this"
  },
  {
    "start": 1898.02,
    "end": 1898.18,
    "text": "okay?"
  },
  {
    "start": 1899.1,
    "end": 1903.4,
    "text": "We've talked a lot about do AI agents behave like humans?"
  },
  {
    "start": 1903.74,
    "end": 1914.36,
    "text": "Just a wild question, but do you think there is something else we from human interaction could learn in regards to helping this specific problem?"
  },
  {
    "start": 1915.2,
    "end": 1919.18,
    "text": "I personally find social science papers very interesting."
  },
  {
    "start": 1919.68,
    "end": 1936.76,
    "text": "And if he focuses on them and studies it can also apply some of the... agent networks to find out how they are communicating or house the information process between them."
  },
  {
    "start": 1937.04,
    "end": 1947.72,
    "text": "And yeah, I think it's really helpful to study human interaction and then maybe find some insights from them about multi-agent systems."
  },
  {
    "start": 1948.16,
    "end": 1954.06,
    "text": "because when we think about that agents are trained on human data right?"
  },
  {
    "start": 1954.66,
    "end": 1958.8,
    "text": "So It's really useful to get some insights in my opinion."
  },
  {
    "start": 1959.3,
    "end": 1962.9,
    "text": "Okay, very interesting thank you for these insights."
  },
  {
    "start": 1963.18,
    "end": 1971.18,
    "text": "it is a topic that as I said sounds quite futuristic but we are already probably at the middle of getting things thrown out."
  },
  {
    "start": 1973.64,
    "end": 1981.86,
    "text": "before i let you go We always have five quick questions so all this will have to do with AI."
  },
  {
    "start": 1983.06,
    "end": 1986.96,
    "text": "So the first one is, what's something that you wish AI could do already?"
  },
  {
    "start": 1987.44,
    "end": 1988.06,
    "text": "Very interesting."
  },
  {
    "start": 1988.22,
    "end": 1995.94,
    "text": "There are a lot of stuff I wanted to do but actually... ...I like physical agents more than they can do some house stuff!"
  },
  {
    "start": 1998.16,
    "end": 1998.9,
    "text": "To be honest."
  },
  {
    "start": 1999.08,
    "end": 2001.12,
    "text": "Okay so you want a housekeeping robot?"
  },
  {
    "start": 2001.36,
    "end": 2002.0,
    "text": "yes yeah"
  },
  {
    "start": 2002.42,
    "end": 2004.76,
    "text": "okay if you find one it works please send me link."
  },
  {
    "start": 2007.16,
    "end": 2012.22,
    "text": "I'm looking forward to the improvements in physical agents part."
  },
  {
    "start": 2012.48,
    "end": 2012.74,
    "text": "Okay,"
  },
  {
    "start": 2013.16,
    "end": 2013.62,
    "text": "i see that."
  },
  {
    "start": 2013.72,
    "end": 2013.84,
    "text": "yeah!"
  },
  {
    "start": 2014.86,
    "end": 2015.42,
    "text": "I understand."
  },
  {
    "start": 2015.66,
    "end": 2019.18,
    "text": "Konjari what is one thing you definitely do not want AI for?"
  },
  {
    "start": 2019.5,
    "end": 2019.98,
    "text": "Thinking?"
  },
  {
    "start": 2021.44,
    "end": 2035.14,
    "text": "Yeah maybe...I don't know if you have experience or after using AI sometimes at some point it to think on behalf of me, I don't want this anymore."
  },
  {
    "start": 2035.44,
    "end": 2042.3,
    "text": "Yeah, i don't wanna lose that creative ideas in my mind and yeah...I don't wants you do the thinking instead of me?"
  },
  {
    "start": 2042.96,
    "end": 2043.08,
    "text": "Yeah!"
  },
  {
    "start": 2043.22,
    "end": 2045.02,
    "text": "I'm interested like all studies coming out."
  },
  {
    "start": 2045.48,
    "end": 2047.3,
    "text": "how is changing our creativity?"
  },
  {
    "start": 2047.38,
    "end": 2047.84,
    "text": "or thinking?"
  },
  {
    "start": 2048.6,
    "end": 2050.359,
    "text": "because there's also a lot happening right now."
  },
  {
    "start": 2051.459,
    "end": 2052.319,
    "text": "Yeah exactly.."
  },
  {
    "start": 2052.359,
    "end": 2053.239,
    "text": "I think its good approach."
  },
  {
    "start": 2055.52,
    "end": 2059.219,
    "text": "What are your skills that humans will always be better at than AI?"
  },
  {
    "start": 2060.08,
    "end": 2061.199,
    "text": "Very interesting question."
  },
  {
    "start": 2062.02,
    "end": 2063.679,
    "text": "Maybe creativity?"
  },
  {
    "start": 2063.759,
    "end": 2070.759,
    "text": "in my opinion, humans will always be more creative than AI agents At least at this point."
  },
  {
    "start": 2070.899,
    "end": 2074.319,
    "text": "I think like that because AI is trained on human data right?"
  },
  {
    "start": 2074.379,
    "end": 2080.319,
    "text": "So... I don't understand how creative they can be Because it's the past data."
  },
  {
    "start": 2080.899,
    "end": 2083.419,
    "text": "But humans would um.. Be more creative."
  },
  {
    "start": 2083.459,
    "end": 2085.139,
    "text": "then these models"
  },
  {
    "start": 2085.52,
    "end": 2089.899,
    "text": "And AI model only working with stuff It has seen before Okay."
  },
  {
    "start": 2091.439,
    "end": 2099.52,
    "text": "Now that we've been comparing AI to humans in this episode, if you could give an AI agent one human trait what would it be?"
  },
  {
    "start": 2100.72,
    "end": 2102.879,
    "text": "It wouldn't be a stubbornness for sure!"
  },
  {
    "start": 2103.08,
    "end": 2103.54,
    "text": "Definitely."
  },
  {
    "start": 2104.52,
    "end": 2104.799,
    "text": "But"
  },
  {
    "start": 2106.379,
    "end": 2111.1,
    "text": "I think you already answered the follow-up question because there will be one trade that you definitely won't get."
  },
  {
    "start": 2111.16,
    "end": 2112.64,
    "text": "so can you have taken that before?"
  },
  {
    "start": 2112.72,
    "end": 2112.879,
    "text": "but"
  },
  {
    "start": 2113.62,
    "end": 2117.04,
    "text": "yeah maybe i would say trustworthiness"
  },
  {
    "start": 2117.66,
    "end": 2119.16,
    "text": "Yeah, something like honesty."
  },
  {
    "start": 2119.72,
    "end": 2126.399,
    "text": "I would like any AI model that I'm attracted to be honest and trustworthy."
  },
  {
    "start": 2126.72,
    "end": 2129.24,
    "text": "Okay well then thanks a lot."
  },
  {
    "start": 2129.54,
    "end": 2135.66,
    "text": "um i hope you will continue down this path of research And have great PhD career at SISPA."
  },
  {
    "start": 2135.72,
    "end": 2136.24,
    "text": "Thank"
  },
  {
    "start": 2136.279,
    "end": 2136.779,
    "text": "You so much."
  },
  {
    "start": 2137.54,
    "end": 2138.939,
    "text": "Great Career in Cybersecurity."
  },
  {
    "start": 2139.02,
    "end": 2140.0,
    "text": "uh In general."
  },
  {
    "start": 2140.14,
    "end": 2140.939,
    "text": "thank you So Much."
  },
  {
    "start": 2141.779,
    "end": 2144.22,
    "text": "This wraps up this episode of CIS Partiliar."
  },
  {
    "start": 2144.379,
    "end": 2148.56,
    "text": "A huge thank you to Sumira for giving us all these insights into multi-agent systems!"
  },
  {
    "start": 2148.979,
    "end": 2158.359,
    "text": "If you have enjoyed the episode, don't forget to follow or subscribe on your favorite podcast app and maybe leave a rating or review – it helps getting our word about their research out even more listeners."
  },
  {
    "start": 2158.62,
    "end": 2159.399,
    "text": "I'm Tobias Ebelzeuser."
  },
  {
    "start": 2159.58,
    "end": 2160.24,
    "text": "thanks for listening."
  },
  {
    "start": 2160.5,
    "end": 2164.279,
    "text": "we'll be back soon with more stories from the world of cybersecurity in AI made simple."
  },
  {
    "start": 2164.52,
    "end": 2166.0,
    "text": "until then take care."
  }
]