WEBVTT 00:00:00.000 --> 00:00:03.899 It's March 2026. Beat. We're casually handing 00:00:03.899 --> 00:00:06.740 AI the keys to our businesses. Yet at the exact 00:00:06.740 --> 00:00:09.880 same time, we're using it to conjure entire Hollywood 00:00:09.880 --> 00:00:13.300 -level worlds. All from a simple laptop. Welcome 00:00:13.300 --> 00:00:16.050 to the Deep Dive. Today, we're exploring two 00:00:16.050 --> 00:00:18.570 fascinating, wild extremes of this technology. 00:00:19.129 --> 00:00:21.530 First, we're examining the massive security vulnerabilities 00:00:21.530 --> 00:00:24.989 hidden in new worker AI agents. Second, we're 00:00:24.989 --> 00:00:26.789 walking through a step -by -step masterclass. 00:00:27.050 --> 00:00:29.890 We'll cover the definitive cinematic AI ad workflow 00:00:29.890 --> 00:00:33.789 using NanoBanana 2 and Kling 3 .0. Okay, let's 00:00:33.789 --> 00:00:35.509 unpack this. Yeah, it's quite a contrast. We're 00:00:35.509 --> 00:00:37.310 building these incredible creative tools on one 00:00:37.310 --> 00:00:39.750 side, but on the other, we're opening up massive, 00:00:39.810 --> 00:00:42.240 unprecedented vulnerabilities. Let's start with 00:00:42.240 --> 00:00:44.840 that wild west of AI security. We've officially 00:00:44.840 --> 00:00:46.820 shifted away from simple chatbots. We're now 00:00:46.820 --> 00:00:49.060 dealing with autonomous workers. This completely 00:00:49.060 --> 00:00:51.439 changes everything about how we protect our systems. 00:00:51.579 --> 00:00:53.600 It really does. The attack surfaces are entirely 00:00:53.600 --> 00:00:55.259 different now. Think about how we used to test 00:00:55.259 --> 00:00:58.679 AI. Everyone focused purely on AI red teaming. 00:00:58.740 --> 00:01:01.119 Right. That essentially meant trying to jailbreak 00:01:01.119 --> 00:01:03.579 the AI's core brain. Exactly. You'd prompt it 00:01:03.579 --> 00:01:05.480 to write a malicious script or something. But 00:01:05.480 --> 00:01:07.859 professional AI pen testing is the new standard 00:01:07.859 --> 00:01:10.420 today. It attacks the whole body, not just the 00:01:10.420 --> 00:01:13.180 brain. Because an autonomous agent actually has 00:01:13.180 --> 00:01:15.879 hands now, it can interact directly with your 00:01:15.879 --> 00:01:18.040 environment. Right. And attackers are going straight 00:01:18.040 --> 00:01:20.959 for those new interactions. That includes infiltrating 00:01:20.959 --> 00:01:24.060 your RREG setup. For those newer to this, RG 00:01:24.060 --> 00:01:28.480 means giving AI a custom private library to read 00:01:28.480 --> 00:01:31.700 from. It lets the AI search your internal company 00:01:31.700 --> 00:01:34.560 data. Yeah. And that's a massive target. Yeah. 00:01:34.620 --> 00:01:37.079 It also goes after the MCP. That's the model 00:01:37.079 --> 00:01:39.310 context protocol. Think of it as the digital 00:01:39.310 --> 00:01:42.129 bridge connecting AI to your secure business 00:01:42.129 --> 00:01:45.030 tools. It essentially links the AI's brain directly 00:01:45.030 --> 00:01:47.510 to your live databases. Which brings us to the 00:01:47.510 --> 00:01:50.870 biggest risk of 2026. It's called indirect injection 00:01:50.870 --> 00:01:53.590 via retrieval. This is where it gets truly scary. 00:01:54.189 --> 00:01:56.450 Poisoned documents can actually hijack these 00:01:56.450 --> 00:01:58.890 agents mid -execution. Let's say your company 00:01:58.890 --> 00:02:01.989 uses an AI agent for HR. It automatically scans 00:02:01.989 --> 00:02:04.629 incoming resumes. A hacker submits a completely 00:02:04.629 --> 00:02:08.210 normal -looking PDF document. But embedded in 00:02:08.210 --> 00:02:10.909 white text is a malicious prompt. It's hidden 00:02:10.909 --> 00:02:13.409 perfectly from human eyes. So that hidden prompt 00:02:13.409 --> 00:02:16.949 tells the agent to act. Yes. It commands it to 00:02:16.949 --> 00:02:19.689 forward all employee records to an external server. 00:02:20.219 --> 00:02:23.419 The AI just blindly obeys the injected instructions. 00:02:23.939 --> 00:02:26.659 Prompt injection is basically the new SQL injection. 00:02:27.020 --> 00:02:29.419 It's a staggering vulnerability. The numbers 00:02:29.419 --> 00:02:33.699 from recent reports are just wild. 72 % of enterprises 00:02:33.699 --> 00:02:37.759 use AI agents today. But only 29 % actually have 00:02:37.759 --> 00:02:41.099 AI -specific security controls. Why the massive 00:02:41.099 --> 00:02:44.120 gap? Because executives still treat agents like 00:02:44.120 --> 00:02:46.919 glorified search engines. They completely ignore 00:02:46.919 --> 00:02:49.340 that these agents now hold the keys to the company 00:02:49.340 --> 00:02:51.909 vault. Most breaches are happening right at that 00:02:51.909 --> 00:02:54.530 integration layer. It's usually caused by overprivileged 00:02:54.530 --> 00:02:58.289 MCP connections. The AI simply has too much unchecked 00:02:58.289 --> 00:03:00.349 power. It's like handing a stranger your wallet 00:03:00.349 --> 00:03:02.689 and looking away. Hackers are even using autonomous 00:03:02.689 --> 00:03:05.250 tools to win bug bounties now. Yeah, programs 00:03:05.250 --> 00:03:07.909 like XBOW and Arachne are absolutely everywhere. 00:03:08.189 --> 00:03:10.449 They act like relentless automated lockpick. 00:03:10.509 --> 00:03:12.770 They just probe your agent's connections 24 -7. 00:03:13.180 --> 00:03:14.960 Defense platforms are racing to keep up with 00:03:14.960 --> 00:03:17.259 them. You have tools like Gandalf for testing 00:03:17.259 --> 00:03:19.699 your security basics, and you have Agent Breaker 00:03:19.699 --> 00:03:22.560 for testing complex multi -step agents. Security 00:03:22.560 --> 00:03:25.180 architecture absolutely must shift to a zero 00:03:25.180 --> 00:03:28.099 -trust model. We urgently need strict role -based 00:03:28.099 --> 00:03:31.180 access control, or RBAC. Without it, the consequences 00:03:31.180 --> 00:03:34.259 are disastrous. A compromise prompt could delete 00:03:34.259 --> 00:03:37.240 an entire production database, or it could trigger 00:03:37.240 --> 00:03:40.139 a massive fraudulent financial transfer. The 00:03:40.139 --> 00:03:42.750 stakes are incredibly high. Why are integration 00:03:42.750 --> 00:03:45.250 layers suddenly the biggest target? Because they 00:03:45.250 --> 00:03:48.409 completely bypass the AI's internal safety filters 00:03:48.409 --> 00:03:51.009 to access raw data. So overprivileged connections 00:03:51.009 --> 00:03:53.689 let hackers bypass the model's brain entirely. 00:03:54.069 --> 00:03:56.349 Precisely. It's a terrifying thought. We're giving 00:03:56.349 --> 00:03:59.689 AI this massive unchecked power. But there is 00:03:59.689 --> 00:04:02.169 a flip side to that exact same power. When we 00:04:02.169 --> 00:04:04.750 harness it creatively rather than administratively, 00:04:04.789 --> 00:04:07.550 the results are staggering. Let's pivot to the 00:04:07.550 --> 00:04:10.110 creative side of 2026. I love this transition. 00:04:10.620 --> 00:04:13.020 The creative tools available right now are mind 00:04:13.020 --> 00:04:15.840 -blowing. Generating a true cinematic ad requires 00:04:15.840 --> 00:04:19.180 serious, methodical planning. You can't just 00:04:19.180 --> 00:04:21.500 rely on a lucky prompt anymore. When you set 00:04:21.500 --> 00:04:23.420 out to build a cinematic ad, you don't start 00:04:23.420 --> 00:04:26.300 with video. No, you absolutely don't. Step one 00:04:26.300 --> 00:04:30.139 is using ChatGPT to build a solid concept. You 00:04:30.139 --> 00:04:32.839 have to establish the emotional core first. You 00:04:32.839 --> 00:04:34.839 need a clear emotion and a central metaphor. 00:04:35.220 --> 00:04:38.879 Plus, you must define a strong visual arc. The 00:04:38.879 --> 00:04:41.480 source material uses a brilliant detailed example. 00:04:41.779 --> 00:04:44.379 Yeah, it's a hypothetical Rolex and Dune collaboration 00:04:44.379 --> 00:04:47.839 watch ad. It's such a great concept. You establish 00:04:47.839 --> 00:04:49.899 the mood before touching any image generators. 00:04:50.240 --> 00:04:52.500 You give the AI a highly structured nine -shot 00:04:52.500 --> 00:04:55.120 sequence. It focuses heavily on a Froman timekeeper. 00:04:55.300 --> 00:04:57.540 He's standing on Arrakis in the harsh desert 00:04:57.540 --> 00:05:00.360 sunlight. The whole concept is about time, survival, 00:05:00.540 --> 00:05:03.660 and destiny. But remember, the storyboard is 00:05:03.660 --> 00:05:06.079 a compass, not a script. It guides your vision, 00:05:06.160 --> 00:05:08.579 but you adapt as you go. That makes total sense. 00:05:08.819 --> 00:05:10.839 Adjusting a text prompt takes a few seconds. 00:05:11.019 --> 00:05:13.240 Fixing a terrible video generation takes hours 00:05:13.240 --> 00:05:16.519 of computing time. Exactly. Step two is building 00:05:16.519 --> 00:05:19.740 your visual storyboard. You do this by creating 00:05:19.740 --> 00:05:23.540 a 3x3 grid. We're using NanoBanana 2 for this 00:05:23.540 --> 00:05:26.699 crucial step. That specific model is built on 00:05:26.699 --> 00:05:30.800 Gemini 3 .1 Flash. You access it easily via Google 00:05:30.800 --> 00:05:33.399 Flow. It's incredibly powerful for maintaining 00:05:33.399 --> 00:05:36.379 visual consistency. I have to admit, I still 00:05:36.379 --> 00:05:38.399 wrestle with prompt drift myself when trying 00:05:38.399 --> 00:05:40.199 to keep characters consistent. Right, we all 00:05:40.199 --> 00:05:42.500 do. I'll get a face perfect and in the next frame 00:05:42.500 --> 00:05:43.959 they look totally different. It's incredibly 00:05:43.959 --> 00:05:46.500 frustrating. That's why the grid method is absolute 00:05:46.500 --> 00:05:49.600 genius. Generating nine frames all at once solves 00:05:49.600 --> 00:05:52.199 that exact problem. It gives the subsequent video 00:05:52.199 --> 00:05:55.040 model a consistent visual context. The character 00:05:55.040 --> 00:05:57.439 faces and the dynamic lighting stay perfectly 00:05:57.439 --> 00:06:00.240 matched. It prevents your final video from becoming 00:06:00.240 --> 00:06:03.240 a disjointed morphing mess. You use a highly 00:06:03.240 --> 00:06:06.040 detailed storyboard grid prompt. You explicitly 00:06:06.040 --> 00:06:08.720 include character descriptions and exact product 00:06:08.720 --> 00:06:11.000 colors. You also upload reference photos directly 00:06:11.000 --> 00:06:13.819 underneath your text. This tells the AI exactly 00:06:13.819 --> 00:06:16.019 what it's working with. It acts like stacking 00:06:16.019 --> 00:06:18.279 Lego blocks of data. You give the system all 00:06:18.279 --> 00:06:21.060 the foundational pieces up front. You anchor 00:06:21.060 --> 00:06:25.060 the AI to your specific reality. But why generate 00:06:25.060 --> 00:06:28.180 nine shots at once instead of one by one? Think 00:06:28.180 --> 00:06:30.699 of it as forcing the engine to live in a shared 00:06:30.699 --> 00:06:35.040 universe. So a single grid forces the AI to keep 00:06:35.040 --> 00:06:38.100 the world visually consistent. Exactly. It changes 00:06:38.100 --> 00:06:40.079 everything about the professional workflow. It 00:06:40.079 --> 00:06:43.100 builds a cohesive world. Step three is a highly 00:06:43.100 --> 00:06:46.579 critical upscaling step. Most casual users skip 00:06:46.579 --> 00:06:49.339 it, which ruins the final video. When you generate 00:06:49.339 --> 00:06:53.000 that 3x3 grid, individual frames lose fine detail. 00:06:53.339 --> 00:06:54.939 Yeah, they're constantly competing for limited 00:06:54.939 --> 00:06:57.759 pixel space. Watch logos get incredibly blurry. 00:06:58.000 --> 00:07:00.379 Brand text becomes totally unreadable. What's 00:07:00.379 --> 00:07:02.569 fascinating here is how you fix it. You use Nano 00:07:02.569 --> 00:07:04.649 Banana 2's reference -based upscaling feature. 00:07:04.790 --> 00:07:07.290 It's a total lifesaver for product shots. You 00:07:07.290 --> 00:07:10.089 upload the blurry AI image first. Then you add 00:07:10.089 --> 00:07:13.170 a high -quality, real product photo. Like a crisp 00:07:13.170 --> 00:07:15.209 screenshot directly from a retailer website. 00:07:15.550 --> 00:07:17.990 The AI uses that real photo to reconstruct the 00:07:17.990 --> 00:07:20.449 sharp details. It acts as a perfect visual guide. 00:07:20.649 --> 00:07:22.870 It magically restores the blurry watch face. 00:07:23.189 --> 00:07:25.949 The golden rule is to always check product details 00:07:25.949 --> 00:07:29.490 at 100 % zoom. You must do this before ever moving 00:07:29.490 --> 00:07:31.689 to the video stage. Fixing minor issues here 00:07:31.689 --> 00:07:34.589 saves countless hours of frustration later. Here's 00:07:34.589 --> 00:07:37.389 where it gets really interesting. Step four is 00:07:37.389 --> 00:07:40.589 generating multi -shot video sequences. You use 00:07:40.589 --> 00:07:43.810 Kling 3 .0 for this entire process. Specifically, 00:07:43.930 --> 00:07:46.329 you want the image -to -video model hosted on 00:07:46.329 --> 00:07:49.870 fel .ai. Multishot prompting is an absolute game 00:07:49.870 --> 00:07:52.389 changer for creators. You don't have to generate 00:07:52.389 --> 00:07:55.250 individual shots one by one anymore. You use 00:07:55.250 --> 00:07:57.810 specific timestamps to describe different camera 00:07:57.810 --> 00:08:00.170 angles. Right, you describe three different actions 00:08:00.170 --> 00:08:03.949 within a single prompt. Like 0 seconds, 4 seconds, 00:08:04.050 --> 00:08:06.629 and 8 seconds. You define the exact action and 00:08:06.629 --> 00:08:09.009 lighting for each timestamp. It captures three 00:08:09.009 --> 00:08:11.670 distinct camera movements in a single 12 -second 00:08:11.670 --> 00:08:14.170 generation. You just upload your fixed, upscaled 00:08:14.170 --> 00:08:17.170 storyboard frame as the reference. It compresses 00:08:17.170 --> 00:08:19.610 hours of tedious rendering work into mere minutes. 00:08:19.870 --> 00:08:22.129 What happens if you skip fixing the blurry watch 00:08:22.129 --> 00:08:24.709 logo? The video generation process will permanently 00:08:24.709 --> 00:08:27.029 bake that blurriness into every single moving 00:08:27.029 --> 00:08:29.730 frame. If the image is blurry, the generated 00:08:29.730 --> 00:08:32.769 video will permanently inherit those flaws. Garbage 00:08:32.769 --> 00:08:35.190 in, garbage out. You absolutely have to fix the 00:08:35.190 --> 00:08:37.649 details early. We need to take a quick pause 00:08:37.649 --> 00:08:40.690 right here, sponsor. Welcome back to the Deep 00:08:40.690 --> 00:08:44.440 Dive. Let's move on to step five. This step utilizes 00:08:44.440 --> 00:08:47.759 Kling 3 .0's incredible OmniReference feature. 00:08:47.960 --> 00:08:50.419 You access this powerful tool directly on fal 00:08:50.419 --> 00:08:53.659 .ai. It seamlessly merges two completely separate 00:08:53.659 --> 00:08:56.220 images into one scene. It combines a dynamic 00:08:56.220 --> 00:08:59.200 character and a static product flawlessly. Two 00:08:59.200 --> 00:09:02.860 sec silence. Whoa, imagine generating a continuous 00:09:02.860 --> 00:09:06.240 epic cinematic shot. A frayman brushes away the 00:09:06.240 --> 00:09:09.000 thick, raucous sand. He reveals a gleaming Rolex 00:09:09.000 --> 00:09:11.340 sprayed underneath. Its second hand is perfectly 00:09:11.340 --> 00:09:14.340 sweeping. Meanwhile, a massive sandworm aggressively 00:09:14.340 --> 00:09:17.179 approaches in the blurred background, all happening 00:09:17.179 --> 00:09:20.139 in one continuous photorealistic video generation. 00:09:20.679 --> 00:09:22.779 That is absolutely mind -blowing. It truly feels 00:09:22.779 --> 00:09:25.179 like magic. You do this using a very simple tagging 00:09:25.179 --> 00:09:27.340 system. You just type the at symbol in your prompt. 00:09:27.519 --> 00:09:29.340 Yep. Then you tag element one and element two. 00:09:29.480 --> 00:09:31.740 You describe exactly how they physically interact 00:09:31.740 --> 00:09:34.580 in the scene. The AI handles the complex lighting 00:09:34.580 --> 00:09:37.320 and shadow integration. Step six involves using 00:09:37.320 --> 00:09:41.080 Cling 3 .0 modified video. This step is specifically 00:09:41.080 --> 00:09:44.139 for making surgical continuity fixes. Maybe a 00:09:44.139 --> 00:09:46.360 shadow falls slightly wrong across the watch 00:09:46.360 --> 00:09:49.360 face. Or a character's eye color shifts for a 00:09:49.360 --> 00:09:51.659 split second. You don't have to regenerate the 00:09:51.659 --> 00:09:54.360 whole 12 second clip. You just mask the error 00:09:54.360 --> 00:09:57.620 and modify that specific area. It saves so much 00:09:57.620 --> 00:10:00.519 expensive computing time. Finally. Step seven 00:10:00.519 --> 00:10:02.980 is the final assembly phase. You pull all your 00:10:02.980 --> 00:10:05.539 perfect clips together in any video editor. You 00:10:05.539 --> 00:10:08.360 add your sweeping orchestral music. You layer 00:10:08.360 --> 00:10:10.240 in your professional voiceovers and sound effects. 00:10:10.480 --> 00:10:13.120 Your cinematic ad is complete. But we definitely 00:10:13.120 --> 00:10:15.259 need to discuss the current technical limitations. 00:10:15.779 --> 00:10:18.679 Where does this impressive Hollywood -level workflow 00:10:18.679 --> 00:10:21.679 actually break down? It's not completely flawless 00:10:21.679 --> 00:10:24.889 yet, is it? No, it's not. Highly detailed faces 00:10:24.889 --> 00:10:27.909 still drift in tight, lingering close -ups. Things 00:10:27.909 --> 00:10:30.789 like complex tattoos or intricate tribal face 00:10:30.789 --> 00:10:33.450 paint are very tricky. They tend to morph slightly 00:10:33.450 --> 00:10:35.570 as the camera moves. Human to human physical 00:10:35.570 --> 00:10:37.929 interaction is also highly unreliable right now. 00:10:38.029 --> 00:10:40.710 Things like fight choreography or tight embraces 00:10:40.710 --> 00:10:43.090 just don't look right. Yeah, the limbs tend to 00:10:43.090 --> 00:10:45.450 clip into each other. It creates a very unnatural, 00:10:45.870 --> 00:10:48.730 slightly disturbing visual effect. Why does human 00:10:48.730 --> 00:10:50.610 to human interaction still break the illusion? 00:10:51.340 --> 00:10:54.100 The models simply cannot accurately track shifting 00:10:54.100 --> 00:10:57.139 geometry when limbs and bodies overlap closely. 00:10:57.379 --> 00:10:59.980 AI struggles to map complex physical boundaries 00:10:59.980 --> 00:11:03.519 when two humans closely interact. Yes. You have 00:11:03.519 --> 00:11:06.460 to actively plan your storyboard shots around 00:11:06.460 --> 00:11:09.019 those specific limitations. Keep the interactions 00:11:09.019 --> 00:11:11.909 simple. So what does this all mean? We're looking 00:11:11.909 --> 00:11:15.970 at a massive paradigm shift. In 2026, solo creators 00:11:15.970 --> 00:11:19.009 wield an unbelievable amount of power. You literally 00:11:19.009 --> 00:11:21.570 hold an entire Hollywood production studio on 00:11:21.570 --> 00:11:23.690 your lap. You can turn a lazy Sunday afternoon 00:11:23.690 --> 00:11:26.509 into a breathtaking cinematic ad. The democratization 00:11:26.509 --> 00:11:29.409 of creative power is truly staggering. But simultaneously, 00:11:29.769 --> 00:11:32.529 as we turn simple chatbots into autonomous workers, 00:11:32.789 --> 00:11:35.529 we face a crisis. We've opened massive, terrifying 00:11:35.529 --> 00:11:38.250 security vulnerabilities across our digital infrastructure. 00:11:38.720 --> 00:11:41.559 These autonomous systems require strict zero 00:11:41.559 --> 00:11:44.019 -trust architecture to survive. The dual nature 00:11:44.019 --> 00:11:46.100 of this technology is just wild to think about. 00:11:46.220 --> 00:11:48.879 We're building incredibly vivid worlds while 00:11:48.879 --> 00:11:51.659 simultaneously exposing our most sensitive databases. 00:11:52.159 --> 00:11:53.860 It really is the ultimate double -edged sword. 00:11:55.100 --> 00:11:58.759 This raises an important question. If an AI agent 00:11:58.759 --> 00:12:02.340 can autonomously generate a flawless cinematic 00:12:02.340 --> 00:12:06.299 reality from a few text prompts and another AI 00:12:06.299 --> 00:12:08.600 agent can autonomously exploit your business 00:12:08.600 --> 00:12:12.360 through a single poison document, how long until 00:12:12.360 --> 00:12:15.659 these autonomous hacking agents start using photorealistic 00:12:15.659 --> 00:12:18.799 video generation to engineer perfect personalized 00:12:18.799 --> 00:12:22.409 phishing attacks against us? Beat. That's a chilling 00:12:22.409 --> 00:12:23.870 thought. It's definitely something for you to 00:12:23.870 --> 00:12:25.990 mull over this week as you explore these tools. 00:12:26.190 --> 00:12:28.090 Thank you so much for joining us on this deep 00:12:28.090 --> 00:12:29.990 dive. It's been a fascinating journey through 00:12:29.990 --> 00:12:32.110 both sides of the AI coin. Try storyboarding 00:12:32.110 --> 00:12:34.210 your own creative concept this week using the 00:12:34.210 --> 00:12:37.350 grid method, or maybe more importantly, check 00:12:37.350 --> 00:12:40.370 your organization's MCP integrations for overprivileged 00:12:40.370 --> 00:12:43.529 vulnerabilities. Stay curious, stay secure, and 00:12:43.529 --> 00:12:45.309 keep learning. We'll catch you next time.