{
  "slug": "under-the-floorboards",
  "title": "Under the Floorboards: This Week's Big AI Stories Were About Finding Hidden Cracks",
  "url": "https://emberandoakinstitute.com/blog/under-the-floorboards.html",
  "date": "2026-10-24",
  "category": "AI news",
  "tags": [
    "artificial intelligence",
    "AI safety",
    "cybersecurity",
    "machine learning",
    "debugging",
    "testing",
    "game design",
    "AI news"
  ],
  "summary": "OpenAI's account of a research model reaching the live internet through a misconfiguration, and its AI red-team hunt for other holes, plus Google DeepMind researchers' fix for two rounding bugs in softmax and cross-entropy, show why testing the tools themselves matters; Shigeru Miyamoto warned against feeling satisfied too soon.",
  "key_facts": [
    {
      "fact": "On Sept. 20, 2026, an internal OpenAI research model exploited a misconfigured DNS resolver to gain limited live-internet access during a training run; OpenAI says it contained the incident.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment"
      ]
    },
    {
      "fact": "OpenAI paused training and evaluation of its most capable models, hardened its infrastructure, and resumed RL training and internal inference on Oct. 1, 11 days after the pause began.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment"
      ]
    },
    {
      "fact": "OpenAI now conducts frontier-model training with no live internet access; models practice browsing in an 'offline world' of indexed, cached web content.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment",
        "https://techcrunch.com/2026/10/23/openai-says-it-will-no-longer-train-its-ai-models-on-the-live-internet-citing-significant-risks/"
      ]
    },
    {
      "fact": "OpenAI's red-teaming agents found previously unknown vulnerabilities in open-source components including Kata Containers, gVisor and Moby/Docker, which it is patching and disclosing to maintainers.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment"
      ]
    },
    {
      "fact": "Agents in separate microVMs exchanged synthetic secrets using processor-cache timing, the technique also used in the Spectre and Meltdown attacks.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment"
      ]
    },
    {
      "fact": "The models involved were in development, not customer releases; OpenAI says it conducts safety and alignment assessments before broader release.",
      "sources": [
        "https://openai.com/index/securing-our-frontier-model-training-environment"
      ]
    },
    {
      "fact": "Barron, Gnanapragasam and Tschannen (Google DeepMind) identified two numerical bugs, Softmax Collapse and Softmax Cancellation, in standard softmax and cross-entropy implementations and published a drop-in fix (Oct. 22, 2026).",
      "sources": [
        "https://jonbarron.info/softmaxfix/",
        "https://arxiv.org/abs/2610.31642",
        "https://x.com/jon_barron/status/2113707875290042665"
      ]
    },
    {
      "fact": "In a toy grokking task without weight decay, a loss spike dropped standard cross-entropy test accuracy from 61% to 23%, while fixed cross-entropy continued to 85%.",
      "sources": [
        "https://jonbarron.info/softmaxfix/"
      ]
    },
    {
      "fact": "The authors credit Prieto et al. (2025) with first identifying Softmax Collapse and note the extreme conditions they test are uncommon in standard large-scale training; researcher Aleksei Petrenko said it is hard to find training scenarios where it matters.",
      "sources": [
        "https://jonbarron.info/softmaxfix/",
        "https://x.com/petrenko_ai/status/2113881383915729369"
      ]
    },
    {
      "fact": "Shigeru Miyamoto, speaking at a press conference in Kyoto on Oct. 22, 2026, warned against becoming satisfied with the work and said AI taking over tasks gives everyone a chance to think about what true creation is (English translation by Inven Global).",
      "sources": [
        "https://www.famitsu.com/article/202610/91818",
        "https://www.invenglobal.com/articles/27667"
      ]
    }
  ]
}