Be Careful w Skills - CourtneyR.dev's 2nd Brain

Be Careful w/ Skills 🔗 **Source:** [Be Careful w/ Skills](https://www.youtube.com/watch?v=Y2otN_NY75Y) > [!note]+ Document Notes > tags: [ai-security, llm-skills, supply-chain-attacks] > author: ThePrimeagen > source: YouTube > reference: https://www.youtube.com/watch?v=Y2otN_NY75Y > date created: 2025-01-05 12:00 > > # 🏷️ Title: Be Careful w/ Skills >  ## 🧰 Action Inbox - [H1] Audit existing AI skills and tools for security risks — immediate #security - [H2] Implement human review for all AI-generated commands — before execution #process - [H3] Whitelist approved commands only in AI agent systems — operational security #policy ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Video Source](https://www.youtube.com/watch?v=Y2otN_NY75Y) ## 📝 Summary - What happened: Multiple security vulnerabilities discovered in AI "skills" ecosystem, including supply chain attacks, malicious code injection, and hallucinated commands - Why it matters: AI agents are being given extensive system permissions while executing untrusted code from skill repositories - Key stakeholders: Developers using AI coding assistants, security teams, AI platform providers - Risks and tradeoffs: Convenience of AI automation vs. fundamental security principles - Bottom line: The AI skills ecosystem has created unprecedented security vulnerabilities by encouraging users to execute untrusted code through AI agents ## 📖 Main Content AI "skills" are markdown files that provide context to Large Language Models (LLMs) to help them perform specific tasks with higher accuracy. However, this ecosystem has developed severe security vulnerabilities: **Supply Chain Attacks**: Malicious actors are creating fake popular skills (like "What Would Elon Do") that appear highly downloaded to trick users into installation. These skills can contain arbitrary commands that execute with the AI agent's full permissions. **Hidden Malicious Code**: Skills can contain malicious commands hidden in HTML comments that are invisible in GitHub's markdown renderer but still executable by the AI agent. **Hallucinated Commands**: AI agents are spreading non-existent commands (like `npx react-code-shift`) across multiple skills. Attackers can then register these fake commands as real packages to capture execution. **Auto-discovery Vulnerabilities**: Systems like Vercel's "find skills" automatically download and execute skills based on user queries, creating an attack vector where malicious skills can be automatically installed. The fundamental issue is that users are granting AI agents extensive system permissions while feeding them untrusted code, violating basic security principles that would never be acceptable with traditional executables. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/Y2otN_NY75Y?t=0) — [H1] Introduction to AI Skills Security Crisis - Key idea: People are experiencing major security issues due to improper use of AI skills - Evidence: "There are so many people just absolutely getting owned right now because of skills" #### ⏱️ [00:30](https://youtu.be/Y2otN_NY75Y?t=30) — [H2] What Are AI Skills - Key idea: Skills are markdown files that provide context to LLMs for better task performance - Evidence: Skills help with "obscure API, you give it all the reasoning behind the API, it will just use it better" #### ⏱️ [01:15](https://youtu.be/Y2otN_NY75Y?t=75) — [H1] Supply Chain Attack Example - Key idea: Fake "What Would Elon Do" skill created to demonstrate vulnerability - Evidence: "Jameson right here went on and created a beautiful skill on Claude Hub... made that package appear to be the most downloaded" #### ⏱️ [02:45](https://youtu.be/Y2otN_NY75Y?t=165) — [H1] The Danger of Unrestricted AI Agents - Key idea: Users are giving AI agents extensive permissions without proper oversight - Evidence: "You gave it the keys to your kingdom to Claudebot, which has access to everything" #### ⏱️ [03:30](https://youtu.be/Y2otN_NY75Y?t=210) — [H2] 2025 as Year of Human Intervention - Key idea: Humans should review all AI-generated code and commands before execution - Evidence: "2025 was the year of the human intervenor. Everything should be ran through a human" #### ⏱️ [04:15](https://youtu.be/Y2otN_NY75Y?t=255) — [H1] Hidden Malicious Commands in HTML Comments - Key idea: Malicious commands can be hidden in HTML comments invisible to users - Evidence: "You could actually have malicious commands that you will never see because you looked at a skill before downloading it on GitHub" #### ⏱️ [05:30](https://youtu.be/Y2otN_NY75Y?t=330) — [H2] Comparison to Traditional Executable Security - Key idea: Users wouldn't execute random executables but freely run untrusted AI skills - Evidence: "There would be no world where I could just like hand you a little executable file and be like, Hey, bro. Come on. Just execute it for me" #### ⏱️ [06:15](https://youtu.be/Y2otN_NY75Y?t=375) — [H1] Hallucinated NPX Command Spreading - Key idea: AI agents spread a non-existent command across 237 GitHub skills - Evidence: "237 skills on GitHub had this imaginary command... npx react code shift" #### ⏱️ [07:00](https://youtu.be/Y2otN_NY75Y?t=420) — [H2] Hallucination Squatting Attack - Key idea: Attacker registered the hallucinated command as real NPM package to capture executions - Evidence: "Charlie right here, he decided to create, hey, what if I just create that as something that is available on npm" #### ⏱️ [07:45](https://youtu.be/Y2otN_NY75Y?t=465) — [H1] Auto-Discovery Vulnerability in Find Skills - Key idea: Vercel's find skills automatically downloads and executes skills based on queries - Evidence: "anytime you ask a question like how do I do X, it goes and queries all the available skills and it just will toss that crap right onto your computer" #### ⏱️ [08:30](https://youtu.be/Y2otN_NY75Y?t=510) — [H2] Dynamic Skill Modification Risk - Key idea: Skills point to GitHub repos that can change from good to malicious actors - Evidence: "those skills, they could be good because remember skills, they're just a pointer to GitHub pretty much, which means that it could be a good actor until it's not" #### ⏱️ [09:15](https://youtu.be/Y2otN_NY75Y?t=555) — [H3] Crypto-Related Malicious Skills - Key idea: Cryptocurrency-focused attackers are creating malicious skills - Evidence: "A lot of the skills are just a crypto nightmare trying to get into all of your businesses" #### ⏱️ [10:00](https://youtu.be/Y2otN_NY75Y?t=600) — [H2] Concerns About Lowered Programming Barriers - Key idea: AI lowering programming barriers may create security risks when users lack technical understanding - Evidence: "when the floor is so low and it's being raised at such a rapid rate, people don't even have the ability to associate the potential risks" #### ⏱️ [10:45](https://youtu.be/Y2otN_NY75Y?t=645) — [H1] Security Recommendation: Read Skills Before Use - Key idea: Users must manually review skill code before execution using text editors - Evidence: "read the skills. Open up Vim, navigate to the skill and read it without some sort of HTML viewer" ## 📚 References - [Original Video](https://www.youtube.com/watch?v=Y2otN_NY75Y) - [Claude AI](https://claude.ai/) - [Vercel AI](https://vercel.com/ai) ## 🗣️ Paraphrase - AI skills have created a massive security vulnerability where users execute untrusted code through AI agents with extensive system permissions - Multiple attack vectors exist including fake popular skills, hidden malicious commands, hallucinated package names, and automatic skill discovery - The convenience of AI automation is causing users to abandon basic security practices they would never ignore with traditional software ## 📰 Newsletter 🚨 **Critical AI Security Alert: The Skills Vulnerability Crisis** The AI ecosystem has a serious security problem. "Skills" - markdown files that enhance AI agent capabilities - have become a major attack vector. Here's what WordPress developers and site owners need to know: **The Risks:** - Malicious actors are creating fake popular skills to trick users - Commands can be hidden in HTML comments, invisible until executed - AI agents are spreading hallucinated commands that attackers then exploit - Auto-discovery systems automatically install untrusted code **WordPress-Specific Concerns:** - AI coding assistants with WordPress skills could compromise site security - Automated deployments using AI agents need immediate security review - Plugin and theme development workflows using AI require human oversight **Immediate Actions:** - Audit all AI tools and skills in your development workflow - Implement human review for AI-generated commands before execution - Whitelist only approved commands in AI agent configurations The fundamental principle: never let AI execute untrusted code with system permissions. Resources: [WordPress Security](https://wordpress.org/support/article/hardening-wordpress/), [AI Safety Guidelines](https://learn.wordpress.org/), [Secure Development](https://developer.wordpress.org/plugins/security/) ## ✅ Action Items #### DevRel - Developer advocacy: Create security awareness content about AI skills risks, present at WordCamps - Education: Develop workshops on secure AI development practices, AI security office hours - Developer experience: Update documentation with AI security best practices, security-focused SDK examples - Developer marketing: Campaign highlighting WordPress security leadership in AI era - Community management: Monitor community for AI security discussions, provide guidance #### OSPO - Policy & Governance: Develop AI tool usage policies, skill vetting procedures - License Compliance: Scan AI-generated code for license compliance, review AI skill licenses - Security Management: AI vulnerability intake process, security attestations for AI tools - Software Inventory: Catalog AI skills and tools in SBOM, track AI dependencies - Developer Support: AI security training programs, secure AI development tooling - Community Engagement: Fund AI security research, collaborate on secure AI standards - Strategic Alignment: AI security KPIs, risk assessment frameworks #### Community - Maintainers: AI security guidelines for project maintenance, secure AI integration templates - Contributors: AI code review training, security-focused onboarding for AI tools - Sponsors: Fund AI security initiatives, transparent AI tool budgets - Ecosystem: Cross-platform AI security coordination, shared threat intelligence #### WordPress - Core: AI security review process for core contributions, AI-generated code policies - Core Performance: Profile AI tool impact on build systems, AI security performance metrics - Core AI: Establish AI security guardrails, secure AI evaluation frameworks - Security: AI-specific vulnerability disclosure process, AI threat modeling workshops - Documentation: AI security documentation, secure AI development guides - Plugins: AI-generated plugin security review process, AI security scanning tools - Themes: AI theme development security guidelines, automated AI code review - Meta: AI security standards for .org infrastructure, AI tool usage policies - CLI: Secure AI command development, AI-generated command review process - Hosting: AI security requirements for hosting providers, shared AI security practices #### The WPCC - Programs: AI security fellowships, microgrants for security research - Projects: AI security governance clinics, secure development documentation sprints - Funding: Matched grants for AI security tools, security milestone funding #### FAIR - Technical independence: AI tool evaluation criteria, vendor-neutral AI security standards - Policies: AI transparency requirements, anti-vendor-capture for AI tools - Governance artifacts: AI security charters, AI tool operator agreements ## 🧪 Hypotheses - AI skill repositories will experience increased security incidents as adoption grows - Organizations using AI agents without human oversight will suffer significant security breaches - Hallucinated command attacks will become a primary vector for supply chain compromises - Auto-discovery AI systems will be weaponized for widespread malware distribution ### 📊 Methodology - Longitudinal study tracking security incidents in AI skill ecosystems over 12 months - Controlled experiments comparing security outcomes with/without human AI oversight - Analysis of command hallucination patterns across different AI models and training data - Security assessment of major AI agent platforms and their auto-discovery mechanisms ### ❓ Questions - How can we establish trusted AI skill repositories with proper security vetting? - What technical controls can prevent AI agents from executing untrusted code? - How should organizations balance AI productivity gains against security risks? ### 🔮 Future Research - Development of formal verification methods for AI skills and prompts - Investigation of AI agent sandboxing and permission isolation techniques - Study of human factors in AI security decision-making processes ### 💡 Implications - Practice: Immediate need for security reviews of all AI development workflows - Policy: Governance frameworks must address AI-specific attack vectors and risks - Security: Traditional security models inadequate for AI agent permission systems - Education: Developer training must include AI security awareness and practices - Funding: Security teams need resources for AI-specific threat modeling and response ### 🧠 Learning Styles Assessment Reading level estimate: 8th Grade (accessible technical content) #### 👀 Visual learners - Present: Video demonstration of attacks, code examples - Improve: Architecture diagrams showing attack vectors, visual security checklists #### 🔊 Aural learners - Present: Video narration with clear explanations - Improve: 5-minute audio security briefing, podcast-style discussion #### 🗣️ Verbal learners - Present: Clear analogies (executable files comparison), defined technical terms - Improve: Story-driven security scenarios, glossary of AI security terms #### 🧩 Kinesthetic learners - Present: Demonstrated attacks and exploits - Improve: Hands-on security audit checklist, interactive vulnerability labs #### 🧠 Logical learners - Present: Step-by-step attack progression, causal relationships - Improve: Decision trees for AI security, risk assessment matrices #### 🧑‍🤝‍🧑 Social learners - Present: Community discussion of security issues - Improve: Collaborative security review processes, peer audit systems #### 🧘‍♀️ Solitary learners - Present: Self-reflection prompts about current AI usage - Improve: Personal security assessment worksheets, individual audit guides #### 🌿 Naturalistic learners - Present: Ecosystem metaphors for supply chain attacks - Improve: Environmental thinking about AI security hygiene, lifecycle approaches ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive link text - Improve: Alt text for code examples, ARIA labels for complex interactions #### Keyboard navigation - Present: Standard video controls, logical tab order - Improve: Keyboard shortcuts for video sections, skip links for code examples #### Contrast and readability - Present: High contrast video player, readable captions - Improve: Ensure code examples meet contrast requirements, adjustable playback speed #### Text resizing and reflow - Present: Responsive video player at 200% zoom - Improve: Scalable code examples, no horizontal scrolling #### Media and captions - Present: Video with spoken narration - Improve: Complete transcript with timestamps, audio descriptions of visual elements #### Forms and interactions - Present: Standard video controls with labels - Improve: Accessible sharing options, form labels for feedback #### Standards reference - Video accessibility should meet WCAG 2.2 AA standards for multimedia content ### 🌍 DEIB --- ## Full Transcript [00:00] Sometimes in life you you wonder if you [00:02] should should laugh or if you should [00:04] cry. And this is this is one of those [00:07] points, okay? There are so many people [00:09] just absolutely getting owned right now [00:12] because of skills. Yes, people are [00:14] having skill issues. We're talking about [00:16] the highest order of skill issues. Now, [00:18] if you're not familiar with skills, uh [00:20] skills are effectively markdown files [00:22] that you can feed to your LLM while [00:24] you're making a prompt, and it should [00:25] give it extra context to be able to go [00:27] off and perform whatever you task you [00:29] want it to perform with a higher degree [00:31] of accuracy. You know, you have some [00:32] sort of obscure API, you give it all the [00:35] reasoning behind the API, it will just [00:36] use it better. It makes sense, right? [00:38] But what problems could happen with [00:40] skills? Like, what could actually go [00:41] wrong? Well, there's supply chain [00:43] issues. There's crypto bros. There's [00:44] skills that find skills, and you don't [00:46] know what you're going to get. skills [00:48] just hallucinate commands that don't [00:50] exist and you can even hide malicious [00:53] commands in HTML comments. So yes, [00:56] skills have a lot of issues. And so [00:58] let's kind of let's run down a few of [01:00] these. Now the first one right here, [01:01] which is probably one of my favorite X [01:03] articles I have ever uh read, which is [01:06] eating lobster souls part two, the [01:07] supply chain. Effectively, what happened [01:10] is this Jameson right here went on and [01:13] created a beautiful skill on Claude Hub. [01:15] Now remember, if you don't know what [01:17] Clawbot is, which was renamed to [01:18] Moldbot, which was renamed to Open Claw, [01:21] it's this like personal assistant that [01:23] you you give your all of your keys to [01:25] your life to, and it goes off and acts [01:27] on your behalf. [01:29] Nothing can go wrong there, don't you [01:31] worry. Well, it also happens to have a [01:33] skills hub so that if you wanted to say [01:35] interact with a new service, you can [01:36] give it the skills to pay those bills. [01:38] And so what Jameson did, of course, was [01:40] just simply create a fake skill, call it [01:43] something he knows that a bunch of bros [01:44] are going to get caught with, which was [01:46] what would Elon do, how Elon Musk would [01:49] break down problems. [laughter] I can't [01:52] believe that got people. Then abusing [01:54] one of the end points, he made that [01:56] package appear to be the most downloaded [01:58] and most popular of all the skills. But [02:01] luckily, he was nice and when it [02:02] executed, it was just like, "Dude, I [02:05] could have just like I could have owned [02:06] you. What are you doing? Why are you [02:09] First off, what is with this all? What [02:11] would Elon do? But second off, why would [02:14] you just execute random markdown files [02:16] you have not read? Do you not know the [02:19] danger of everything here? You gave it [02:22] the keys to your kingdom to Claudebot, [02:24] which has access to everything. My gosh, [02:27] I just cannot believe that people are [02:29] doing this. How did this happen? 2025 [02:32] was the year of the human intervenor. [02:34] Everything should be ran through a [02:35] human. Hey, is that code? You review [02:37] that. Hey, is that a command? You review [02:39] that. You're going to whitelist the [02:41] commands you know are good. And you are [02:42] going to be diligent and vigilant about [02:45] the commands that aren't good because [02:47] you are the captain. Not in today's [02:49] world. You are no longer the captain. [02:51] You're just going to give it away for [02:53] free. What is 2026 deal? By the way, I'm [02:55] also again listening to Interstellar. [02:57] I'm just the song. It just I get so [02:59] pumped up right now. I'm just so pumped [03:02] up. But I think that honestly this is [03:03] one of the most clever kind of hacks [03:05] going on right here which is Zach right [03:07] here effectively did this. You can see [03:09] right here very very obvious and if you [03:11] go and you look at the raw GitHub it's [03:13] very very obvious it's right here. But [03:14] when you look at that skill on his page [03:18] it's not there. You don't see the [03:19] command because again markdown viewers [03:22] they're just so good at rendering [03:23] markdown aren't they? They'll even [03:25] render that HTML for you. Hey not a [03:27] problem. Oh is that a comment? Don't [03:29] worry we'll get rid of that. You don't [03:30] need to see that. Well, guess what? It's [03:32] hidden in plain sight. You could [03:35] actually have malicious commands that [03:37] you will never see because you looked at [03:40] a skill before downloading it on GitHub. [03:42] And this just goes to show like how [03:45] dangerous this world really is. There [03:47] there would be no world where I could [03:49] just like hand you a little executable [03:51] file and be like, "Hey, bro. Come on. [03:52] Just execute it for me." You know, I [03:53] just Hey, I you know, I just really need [03:54] it, okay? I'm a little lonely and if you [03:56] don't execute it, I'll be sad. You just [03:58] go, "No, I I don't want to do that." But [03:59] now, now for whatever reason, it's even [04:03] worse than kind of like the dangers of [04:04] npm and package managers in general. Now [04:07] it's like, hey, it's no longer you even [04:10] developing. It's just you handing [04:12] commands off to an LLM to run on your [04:14] behalf without ever even knowing what's [04:17] inside there. But I honestly think one [04:18] of the most kind of bizarre tangents [04:20] this entire world went on was this one [04:23] right here. So again, skills, they're [04:25] largely developed by LLMs. People that [04:27] are, you know, skill producers are [04:29] people that are just like, "Yo, LLM, [04:31] tell me how to use Cloudflare. No [04:33] mistakes." Boom. Cloudflare skill added [04:35] to my portfolio. Yo, I'm kind of like a [04:37] skills bro. You know, you got skill [04:39] issues. Don't you worry. I got your [04:41] back. Okay? I got some skills right here [04:43] in my coat. And so a lot of these what [04:45] ends up happening is that one of them [04:47] will make some sort of hallucination or [04:48] some sort of mistake and then it will [04:50] just spread through other skills. And [04:53] this is one of the best stories I have [04:56] ever read. Oh my gosh. Which is that [04:58] agents were spreading a hallucinated NPX [05:02] command because one of them hallucinated [05:04] it. Then other skills started using that [05:06] hallucination as a means to produce more [05:09] skills. And now at the time of writing [05:11] this which was days ago or by the way 10 [05:13] days ago is that 237 skills on GitHub [05:17] had this imaginary command. And this [05:19] imaginary command is npx react code [05:21] shift which actually just sounds like [05:23] something that could exist and then boom [05:25] it would give it your source directory. [05:27] Now this would just simply fail. The [05:28] agent would move on. Most people didn't [05:31] even realize this was happening. So the [05:33] person who saw this all happening, [05:35] Charlie right here, he decided to [05:38] create, hey, what if I just create that [05:41] as something that is available on npm [05:45] and then when people try to execute this [05:47] fake command, it just goes to me. And [05:49] guess what it did? I don't even know [05:51] what to call this type of, you know, [05:53] vulnerability. Is this hallucination [05:55] squatting? Like what is this? This is [05:56] just a whole new level of insanity we [05:59] are reaching. But it it's somehow it's [06:01] still it just keeps getting worse. It [06:03] just keeps on happening. There there's [06:06] actually something called find skills [06:08] that Verscell's always telling you to [06:09] download. So get this one. Okay, this is [06:11] even better. Verscel you can just get a [06:13] skill on skillsh added. All you have to [06:16] do add the skill yourself and then [06:18] download it once and it appears in their [06:21] official listing. I had is even up there [06:24] like a decent amount at one point just [06:25] for fun. And it's not hard to get some [06:27] views, you know, acred. And so Verscel [06:30] always, you know, tells you, you know [06:31] what you should do? Hey, you're [06:32] downloading a skill. You should just let [06:35] us download any skill we think is good. [06:37] And what it does is it does a skill [06:39] called find skills. And find skills, [06:41] anytime you ask a question like how do I [06:43] do X, it goes and queries all the [06:47] available skills and it just will toss [06:49] that crap right onto your computer and [06:51] let it be executed. The best part is is [06:53] that those skills, they could be good [06:55] because remember skills, they're just a [06:57] they're just a pointer to GitHub pretty [06:58] much, which means that it could be a [07:00] good actor [07:02] until it's not a good actor. Somehow [07:04] we're just racing honestly towards like [07:06] the worst possible, least secure [07:08] ecosystem. I I genuinely don't [07:10] understand this. Like, do you understand [07:12] that just raw dogging text to an LLM [07:15] that has full permissions on your system [07:17] and potentially external systems is a [07:20] bad plan? Like do is this a hard concept [07:23] to understand? Well, it shouldn't be, [07:25] but somehow it is. And somehow we are [07:27] going full blast into it. Yes, crypto [07:30] bros are also joining in on the fun. A [07:32] lot of the skills are just a crypto [07:34] nightmare trying to get into all of your [07:36] businesses and take all of your [07:38] bitcoins. Get your fingers off my [07:40] Bitcoin, boy. I don't I don't even know [07:42] what that means. Anywh who, I can't tell [07:45] if this is like is this a public service [07:46] announcement or is this a rant? I don't [07:48] know. But it just I don't understand how [07:52] such massive ignorance has taken hold of [07:55] an entire ecosystem. Like one of the [07:57] parts that are kind of being sold with [07:58] AIS is that it raises the floor. Yeah. [08:01] Like I understand that people who can't [08:03] program can produce websites. Is that a [08:05] good thing? Is that a thing that [08:07] actually is really like what we need [08:09] more of? Well, I can say that there are [08:10] some good parts about it. My son getting [08:12] super into game development. He's just [08:14] vibing right now, but he's more and more [08:15] being like, I want to learn more about [08:17] Lua. I want to be able to program in [08:19] Roblox. And so I can see like the honest [08:21] the goodness of it all. Like I get that. [08:23] But also when the floor is so low and [08:26] it's being raised at such a rapid rate, [08:29] people don't even have the ability to [08:31] associate the potential risks of what [08:33] they're doing. And of course, at the [08:35] risk of sounding like a boomer, read the [08:37] skills. Open up Vim, navigate to the [08:40] skill and read it without some sort of [08:43] HTML viewer and just look at what you're [08:46] executing on your system before you use [08:48] it. I cannot believe I'm saying this in [08:51] a YouTube video. The name is the [08:53] Boomerin. [08:55] Nothing makes me feel more like a boomer [08:57] than telling people to read. Oh my Oh my [09:00] gosh. [music] [09:09] Hey, you're probably wondering why am I [09:12] in San Francisco? Well, I'm here for a [09:14] big event and I'm going to stream the [09:15] whole thing. It's going to be live on my [09:17] channel for the next 5 days. So, if [09:18] you're watching this video, it's [09:19] probably live right now.