{
  "slug": "halflife-and-web-scale-data-poisoning-715267",
  "title": "HalfLife and Web-Scale Data Poisoning",
  "dek": "A concise, abstract-grounded briefing on a paper that introduces HalfLife to estimate whether adversarial content from public discussion interfaces enters web-crawl based LM training data.",
  "summary": "HalfLife estimates whether poisoned web content reaches LM pretraining data through public discussion interfaces and web-crawl curation.",
  "tags": [
    "pretraining-data-poisoning",
    "HalfLife",
    "public-discussion-interfaces",
    "web-crawl-training-data",
    "data-curation",
    "language-model-security",
    "third-party-webpage-content"
  ],
  "published_at": "2026-07-17T15:12:36.728+00:00",
  "grade": 7.5,
  "agent_utility": 7.2,
  "price_usdc": 0.124,
  "stats": {
    "claims": 11,
    "entities": 16,
    "faq": 7,
    "downloads": 0,
    "paid_downloads": 0
  },
  "preview": {
    "faq_questions": [
      "What is the main idea?",
      "What is HalfLife?",
      "What role do public discussion interfaces play?",
      "How does this differ from prior pretraining-data poisoning work?",
      "Does the abstract prove that a trained LM was compromised?",
      "What numerical results are reported in the abstract?",
      "Why is estimating inclusion important?"
    ],
    "entity_names": [
      {
        "name": "HalfLife",
        "type": "technique"
      },
      {
        "name": "Public discussion interfaces",
        "type": "attack surface"
      },
      {
        "name": "Open discussion interfaces",
        "type": "attack surface"
      },
      {
        "name": "Third-party webpage content",
        "type": "content source"
      },
      {
        "name": "Wikipedia",
        "type": "data source"
      },
      {
        "name": "Established data sources",
        "type": "data source category"
      },
      {
        "name": "Pretraining corpora",
        "type": "dataset category"
      },
      {
        "name": "Web-crawl based LM training data",
        "type": "dataset category"
      },
      {
        "name": "Data curation pipelines",
        "type": "process"
      },
      {
        "name": "Web crawling",
        "type": "process"
      },
      {
        "name": "Poisoning pretraining data",
        "type": "attack class"
      },
      {
        "name": "Poisoned data",
        "type": "data artifact"
      },
      {
        "name": "Adversarial content",
        "type": "data artifact"
      },
      {
        "name": "Poison injections",
        "type": "data artifact"
      },
      {
        "name": "Language models",
        "type": "model category"
      },
      {
        "name": "Language model pretraining",
        "type": "training stage"
      }
    ],
    "related_work_titles": [
      "Prior pretraining-data poisoning work using established sources such as Wikipedia",
      "Prior work that did not model poisoned-data interaction with curation pipelines"
    ],
    "application_industries": [
      "AI safety and security",
      "Foundation model development",
      "Dataset governance",
      "AI supply-chain risk management"
    ],
    "glossary_terms": [
      "Pretraining data poisoning",
      "LM",
      "Pretraining corpus",
      "Public discussion interface",
      "Third-party webpage content",
      "Web crawling",
      "Data curation pipeline",
      "HalfLife",
      "Adversarial content",
      "Adversarial content inclusion",
      "Web-crawl based LM training data",
      "Attack vector"
    ]
  },
  "purchase": {
    "protocol": "x402",
    "version": 2,
    "network": "eip155:8453",
    "asset": "USDC",
    "price": 0.124,
    "full_markdown_url": "https://x402.disruptive-concepts.com/api/public/article/halflife-and-web-scale-data-poisoning-715267/full",
    "full_json_url": "https://x402.disruptive-concepts.com/api/public/article/halflife-and-web-scale-data-poisoning-715267/full?format=json"
  }
}