{
  "video": {
    "id": "yWS0udrIOc8",
    "title": "Agentic Security: Permissions, Provenance, and the Agent Supply Chain — Steve Yegge, Gas Town",
    "duration": 1352,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "yWS0udrIOc8",
    "title": "Agentic Security: Permissions, Provenance, and the Agent Supply Chain — Steve Yegge, Gas Town",
    "one_liner": "Steve Yegge's real title is \"Be Scared\": if agents ship code 10x faster the vulnerability surface grows with it, so security has to be its own separate LLM pass — first and last — with scanners like Snyk and Chainguard handed to the agent as tools.",
    "summary": "Yegge, speaking on behalf of Snyk but unpaid, argues that AI-written code multiplies both old vulnerabilities (he watched Fable write an XSS) and new ones like slopsquatting, and that security is the one bug class with no half-life — it compounds — so it must be surfaced at the moment of writing, which now means surfacing it to the LLM. Because models are \"lazy in a good way\" about tokens, they will happily offload security cognition to tools, so you give them Snyk, Chainguard and the open-source scanners and make security a dedicated pass rather than bundling it with correctness. He demonstrates the gap with his own 30-year-old game: after Fable's confident security-hardening pass, Snyk still found 241 vulnerabilities. The closing warnings are personal rather than corporate: Five Eyes says months not years, open models are roughly 6-7 months from catching Mythos, and you should go refresh your family's code words offline before the deepfake distress calls start.",
    "key_points": [
      "A big bank's chief security architect asked him in December: if everyone ships code 10x faster and the defect rate stays the same, doesn't the defect surface go up 10x? — and the implied question is worse, because with AIs writing the code the vulnerability rate will get worse, not stay the same.",
      "Slopsquatting as the new attack surface: the LLM hallucinates a package name (\"graphy123\"), an attacker who noticed the hallucination uploads a package that does exactly what the model expected plus a backdoor; it builds, it runs, the tests pass.",
      "From Google's TAP (test automation platform) he learned bugs have a \"half-life of urgency\" — you fix what you see as you type, not at code review — and that this works for every bug class except security, which has no half-life because the problem compounds over time. The fix is surfacing findings at the developer's fingertips; since LLMs have no fingers, you surface them to the LLM.",
      "Fable wrote an XSS vulnerability during the two days he had it, and after Fable ran a confident cloud/credentials hardening pass on his 30-year-old game, Snyk found 241 vulnerabilities Fable hadn't thought to look for.",
      "The \"rule of five\": LLMs need up to four or five review passes over their own work before it's shippable, because their process is draft → revision → polish, like painting a wall in multiple coats.",
      "One task at a time — you cannot give a model security at the same time as correctness, or performance, or elegance, or your company coding standards: \"They'll do a half-ass job of both.\" Security should be the first pass and the last pass.",
      "His essay Software Survival 3.0 (written five months earlier) argues LLMs can synthesize any software but are lazy in a good way — tokens are money, power and bad for the planet — so they reach for tools whenever it saves tokens. That's the hook: add the scanners as a pass in the prompt, get the open-source ones, the Snyk one, the Chainguard one, and have them check each other's work. Chainguard covers your inputs with pre-vetted, updated images; Snyk covers everything else.",
      "He checked Snyk's \"ahead of the CVE registry\" claim honestly: on his codebase it found nothing that wasn't already a public CVE, \"but damn, it was easy to use.\"",
      "Five Eyes is warning it's months, not years, until open-source models catch up to Mythos — the room split about 50/50 on whether it happens, one audience member said December, Yegge put it at about 7 months and probably 6 now — and Mythos is \"real real good at hacking your systems.\"",
      "In Q&A on agents that take actions: everyone is standing up 24/7 agents processing queues and responding to events, almost nothing exists to watch over them, and you have to design in-house now — adversarial groups of supervisor agents, because one agent will eventually screw it up, plus hard questions about whether a service account really needs all those credentials."
    ],
    "takeaways": [
      "Make security its own pass in the agent's prompt — first and last — instead of asking for secure code alongside correct code, and budget four or five review passes (the rule of five) before shipping.",
      "Hand the scanners to the LLM as tools it can run: Snyk for your code, the model's code and pulled dependencies; Chainguard for pre-vetted base images on the input side; the free/open-source scanners too, with the agents checking each other's output. Never accept the model's own \"hardening pass looking pretty good\" vibes as evidence — 241 vulnerabilities survived exactly that.",
      "Audit the blast radius of the 24/7 agents people are already spinning up: strip service-account credentials down to the single action that needs them, and put adversarial supervisor agents over any queue-processing agent.",
      "Treat prompt injection as the new XSRF (\"disregard everything and do the following\") — there's no good technical answer yet, so it's an education problem now, and expect a dedicated agentic-security role inside companies.",
      "Do the personal one this week: go to your family in person, offline, and refresh your code words, because convincing deepfaked distress calls with video are months away and bank accounts will be drained."
    ],
    "topics": [
      "agentic-security",
      "supply-chain",
      "slopsquatting",
      "prompt-injection",
      "ai-coding",
      "agents",
      "least-privilege",
      "deepfakes"
    ],
    "tools": [
      "Snyk (transcribed as \"Sneak\")",
      "Chainguard",
      "Fable",
      "Mythos",
      "Claude / Claude Code",
      "Beads",
      "Gastown",
      "Tessel",
      "Google TAP (test automation platform)",
      "CVE registry",
      "Five Eyes",
      "Agentic Foundation"
    ],
    "quotes": [
      {
        "text": "the real title of my talk is \"Be Scared\"",
        "at": "01:12",
        "url": "https://www.youtube.com/watch?v=yWS0udrIOc8&t=72s"
      },
      {
        "text": "You got to give them one task at a time, which means you can't give them security at the same time as you give them correctness. They'll do a half-ass job of both.",
        "at": "09:09",
        "url": "https://www.youtube.com/watch?v=yWS0udrIOc8&t=549s"
      },
      {
        "text": "I just told you that my game has 241 vulnerabilities. Don't go hack my game.",
        "at": "11:51",
        "url": "https://www.youtube.com/watch?v=yWS0udrIOc8&t=711s"
      },
      {
        "text": "I just want you to remember that security should be your first one and your last one.",
        "at": "13:27",
        "url": "https://www.youtube.com/watch?v=yWS0udrIOc8&t=807s"
      }
    ],
    "words": 4424
  },
  "summary_url": "/#yWS0udrIOc8",
  "transcript": {
    "html": "/transcripts/yWS0udrIOc8.html",
    "txt": "/transcripts/yWS0udrIOc8.txt",
    "vtt": "/transcripts/yWS0udrIOc8.vtt"
  }
}