{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 183,
  "pageCount": 213,
  "prevPage": 182,
  "nextPage": 184,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "balanced",
  "nDataPoints": 0,
  "notes": "The slide uses a specific example of an LLM agent performing SQL injection to illustrate the threat.",
  "elementsJson": [
    "bullet_list",
    "paragraph",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/183",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-183",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "A community of red teamers (led by the anonymous Pliny the Prompter) have managed the defenses outlined on the previous slide, with GPT-4o mini’s Instruction Hierarchy being compromised within hours.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d1f-5ec7c5d51730",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.3,
        "x": 0.65,
        "y": 0.45
      },
      "kind": "image",
      "text": "Figure 1. Schematic of using autonomous LLM agents to hack websites.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "58928c24-9c07-456b-94cd-e40efbe0e2f1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.6,
        "w": 0.6,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "A community of red teamers (led by the anonymous Pliny the Prompter) have managed the defenses outlined on the previous slide, with GPT-4o mini's Instruction Hierarchy being compromised within hours.\nWhile much of this work is done by ethically-motivated groups, the UK's AI Safety Institute has expressed alarm at how models from leading labs comply with harmful requests “under relatively simple attacks”.\nAlthough jailbreak attacks are mostly harmless, DeepKeep, an Israeli cybersecurity start-up, made Llama 2 reveal sensitive personal data.\nMeanwhile, a team at UIUC has shown that GPT-4's ability to leverage tool use and long context means it can hack websites by performing tasks like SQL injections without human feedback. With the right context, it can also exploit one-day vulnerabilities.\nOther research has illustrated the vulnerability of multi-agent environments to 'infectious attacks', where single agents are jailbroken, before contaminating the others.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "79aea53d-78ff-4fbf-aad3-c14d57e7338a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.5,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "...but they’re unable to keep up with the red team",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "92dae91b-9d1b-4574-b655-450611b851d1",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Callback",
      "slug": "callback",
      "agent": "Storyteller",
      "layer": "loop",
      "matchId": "019dd95a-1055-74e0-a95a-9abbc4ca77a6",
      "evidence": "'...but' explicitly chains to prior slide",
      "confidence": 75
    },
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a95a-9788ee4ba24c",
      "evidence": "Title '...but they're unable to keep up with the red team'",
      "confidence": 85
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 204,
      "from": 154,
      "beatId": "019dd95a-0682-776c-8e35-605159a069e6",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore consequences and risks",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 204,
      "from": 175,
      "beatId": "019dd95a-0682-776c-8e35-74f05e198faf",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Return",
      "beatSlug": "voyage-return-the-return",
      "evidence": "Safety section returns to risks/governance debate",
      "position": 4,
      "confidence": 55,
      "parentBeatName": "Resolution",
      "parentBeatSlug": "resolution"
    }
  ],
  "loops": [
    {
      "to": 186,
      "from": 182,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-76aa50b143e8",
      "evidence": "Attack surface, Pliny red team, LLM-on-LLM patching, adversarial robustness, stealth attacks",
      "position": 22,
      "objective": "Show jailbreaking arms race across attack surfaces",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}