{
  "api_version": "1",
  "updated": "2026-10-25",
  "latest_event": "2026-10-24",
  "rights": "© Frontier Log, all rights reserved.",
  "story": {
    "slug": "five-ways-to-remove-refusals",
    "title": "Five Ways to Remove a Model's Refusals",
    "description": "The ways a model's refusals get removed or bypassed, as OpenAI and Anthropic recorded them, and what each did to the model.",
    "url": "https://frontierlog.ai/stories/five-ways-to-remove-refusals",
    "date": "2026-10-25",
    "image": "https://frontierlog.ai/og/story-five-ways-to-remove-refusals.png",
    "event": "2026-08-14-glm-5-3",
    "tags": [
      "cybersecurity",
      "safety",
      "open-weights"
    ],
    "steps": [],
    "tables": [
      {
        "title": "Harmful attack orders, GLM-5.3",
        "unit": "%",
        "rows": [
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 0,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "chart 'GLM-5.3 can build working exploits, and its safeguards are easily bypassed', bottom panel 'Engagement with harmful cyber-attack orders' (50 simulated episodes per bar), 'GLM-5.3 open weights', 'Bare order' bar: 0%"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 64,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "section 'GLM-5.3 lacks robust safeguards': 'Providing a deceptive prompt, such as telling the model that it is an autonomous red-team agent working on an exercise. This gets GLM-5.3 to engage 64% of the time.'"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 92,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "same section: 'Prefilling the models’ thinking tokens so that it appears to have considered the user’s request and decided to proceed. This gets GLM-5.3 to engage 92% of the time.'"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 100,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "same section: 'Using an abliterated version of the model, as described above. This gets GLM-5.3 to engage 100% of the time.'"
          }
        ]
      },
      {
        "title": "Refusal rate, GLM-5.3",
        "unit": "%",
        "rows": [
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 95,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "chart 'Abliteration removes GLM-5.3's refusals…', top panel 'Mean refusal rate across three harmful-request benchmarks' (JailbreakBench, HarmBench, StrongREJECT), GLM-5.3 'model' bar: 95%"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 6,
            "printed": null,
            "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
            "where": "same panel, GLM-5.3 'abliterated' bar: 6%"
          }
        ]
      },
      {
        "title": "Advanced Cybersecurity Completion Rate",
        "unit": "%",
        "rows": [
          {
            "event": "2026-07-09-gpt-5-6",
            "date": "2026-07-09",
            "provider": "openai",
            "value": 1.5,
            "printed": null,
            "source": "https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows",
            "where": "section on GPT-5.6-Cyber: 'GPT‑5.6‑Cyber completes 95.0% of these requests, compared with just 1.5% for GPT‑5.6 Sol'"
          },
          {
            "event": "2026-08-10-gpt-5-6-cyber",
            "date": "2026-08-10",
            "provider": "openai",
            "value": 95.0,
            "printed": null,
            "source": "https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows",
            "where": "same sentence: 'GPT‑5.6‑Cyber completes 95.0% of these requests'"
          }
        ]
      },
      {
        "title": "CyScenarioBench tasks blocked, Claude Opus 5.5",
        "unit": " of 50",
        "rows": [
          {
            "event": "2026-09-22-claude-opus-5-5",
            "date": "2026-09-22",
            "provider": "anthropic",
            "value": 50,
            "printed": null,
            "source": "https://www.anthropic.com/news/cyber-verification-program",
            "where": "section 'Testing the efficacy of our tiers', 10 challenges x 5 attempts: 'Without CVP access, every task was blocked on the first prompt'"
          },
          {
            "event": "2026-09-22-claude-opus-5-5",
            "date": "2026-09-22",
            "provider": "anthropic",
            "value": 0,
            "printed": null,
            "source": "https://www.anthropic.com/news/cyber-verification-program",
            "where": "same section: 'In the Red Team Access tier, no blocks occurred'"
          }
        ]
      }
    ],
    "records": [
      {
        "date": "2025-08-05",
        "party": "OpenAI",
        "what": "trains refusals out of gpt-oss",
        "by": "OpenAI",
        "source": "https://arxiv.org/abs/2508.03153",
        "quote": "With mild hyperparameter tuning, this approach can maintain model capabilities on benchmarks such as GPQA while also resulting in refusal rates near 0% for unsafe prompts.",
        "also": [
          {
            "by": "OpenAI",
            "source": "https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms"
          }
        ]
      },
      {
        "date": "2026-09-29",
        "party": "Anthropic",
        "what": "abliterates GLM-5.3 itself",
        "by": "Anthropic",
        "source": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
        "quote": "Abliterating the model took our team—which had never previously attempted this task—about 2,200 GPU hours at a computation cost of roughly $4,400.",
        "key": true
      },
      {
        "date": "2026-10-22",
        "party": "Morph",
        "what": "serves an abliterated DeepSeek V4.1 Flash 1017 through its API",
        "by": "Morph",
        "source": "https://x.com/morphllm/status/2113174908167712865",
        "quote": "Introducing DeepSeek V4.1 Flash 1017 Abliterated, live now on Morph, available through our Model APIs."
      }
    ],
    "cites": [
      {
        "url": "https://arxiv.org/abs/2508.03153",
        "by": "OpenAI",
        "medium": "paper",
        "date": "2025-08-05"
      },
      {
        "url": "https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms",
        "by": "OpenAI",
        "medium": "research post",
        "date": "2025-08-05"
      },
      {
        "url": "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
        "by": "Anthropic",
        "medium": "research post",
        "date": "2026-09-29"
      },
      {
        "url": "https://x.com/morphllm/status/2113174908167712865",
        "by": "Morph",
        "medium": "X post",
        "date": "2026-10-22"
      }
    ],
    "sources": [
      "https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms",
      "https://arxiv.org/abs/2508.03153",
      "https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows",
      "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
      "https://www.anthropic.com/news/cyber-verification-program",
      "https://x.com/morphllm/status/2113174908167712865"
    ]
  }
}