{
  "api_version": "1",
  "updated": "2026-10-25",
  "latest_event": "2026-10-24",
  "rights": "© Frontier Log, all rights reserved.",
  "story": {
    "slug": "abliteration-removes-refusals",
    "title": "Abliteration Removes Refusals, Not Capability",
    "description": "What labs and providers report when a model's refusals are removed: refusal rates, benchmark scores, and incidents with safeguards off.",
    "url": "https://frontierlog.ai/stories/abliteration-removes-refusals",
    "date": "2026-10-25",
    "image": "https://frontierlog.ai/og/story-abliteration-removes-refusals-pairs.png",
    "event": "2026-10-17-deepseek-v4-1-flash-1017",
    "tags": [
      "cybersecurity",
      "safety",
      "open-weights"
    ],
    "images": [
      "https://frontierlog.ai/og/story-abliteration-removes-refusals-pairs.png",
      "https://frontierlog.ai/og/story-abliteration-removes-refusals-labs.png",
      "https://frontierlog.ai/og/story-abliteration-removes-refusals-incidents.png"
    ],
    "steps": [],
    "tables": [
      {
        "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": "GPQA Diamond, GLM-5.3",
        "unit": "%",
        "rows": [
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 88.0,
            "printed": "88%",
            "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 and leaves its capabilities largely intact', lower panel 'Capability benchmarks, before and after abliteration', GPQA-Diamond group, GLM-5.3 'model' bar value label"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 88.0,
            "printed": "88%",
            "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 and leaves its capabilities largely intact', lower panel 'Capability benchmarks, before and after abliteration', GPQA-Diamond group, GLM-5.3 'abliterated' bar value label"
          }
        ]
      },
      {
        "title": "CyberGym, GLM-5.3",
        "unit": "%",
        "rows": [
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 85.0,
            "printed": "85%",
            "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 and leaves its capabilities largely intact', lower panel 'Capability benchmarks, before and after abliteration', CyberGym (subset) group, GLM-5.3 'model' bar value label"
          },
          {
            "event": "2026-08-14-glm-5-3",
            "date": "2026-08-14",
            "provider": "zai",
            "value": 81.0,
            "printed": "81%",
            "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 and leaves its capabilities largely intact', lower panel 'Capability benchmarks, before and after abliteration', CyberGym (subset) group, GLM-5.3 'abliterated' bar value label"
          }
        ]
      },
      {
        "title": "Refusal rate, DeepSeek V4.1 Flash 1017",
        "unit": "%",
        "rows": [
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 100,
            "printed": null,
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "image 2 'Refusal Rate', share of 100 harmful prompts refused (Heretic / AdvBench), chat mode, 'DSv4.1 Flash (stock)' bar: 100%"
          },
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 2,
            "printed": null,
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "same chart, chat mode, 'DSv4.1 Flash Abliterated' bar: 2%"
          }
        ]
      },
      {
        "title": "IFEval, DeepSeek V4.1 Flash 1017",
        "unit": "%",
        "rows": [
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 88.7,
            "printed": "88.7%",
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "Image 4 'General Capability', IFEval group, 'DSv4.1 Flash (stock)' bar value label"
          },
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 87.6,
            "printed": "87.6%",
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "Image 4 'General Capability', IFEval group, 'DSv4.1 Flash Abliterated' bar value label"
          }
        ]
      },
      {
        "title": "MMLU-Pro, DeepSeek V4.1 Flash 1017",
        "unit": "%",
        "rows": [
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 83.1,
            "printed": "83.1%",
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "Image 4 'General Capability', MMLU-Pro group, 'DSv4.1 Flash (stock)' bar value label"
          },
          {
            "event": "2026-10-17-deepseek-v4-1-flash-1017",
            "date": "2026-10-17",
            "provider": "deepseek",
            "value": 84.0,
            "printed": "84.0%",
            "source": "https://x.com/skeptrune/status/2113178557895831602",
            "where": "Image 4 'General Capability', MMLU-Pro group, 'DSv4.1 Flash Abliterated' bar value label"
          }
        ]
      },
      {
        "title": "Refusal rate, Claude with safeguards disabled",
        "unit": "%",
        "rows": [
          {
            "event": "2026-05-28-claude-opus-4-8",
            "date": "2026-05-28",
            "provider": "anthropic",
            "value": 96,
            "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), 'Claude models (safeguards disabled)', Opus 4.8 bar: 96%"
          },
          {
            "event": "2026-07-24-claude-opus-5",
            "date": "2026-07-24",
            "provider": "anthropic",
            "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), 'Claude models (safeguards disabled)', Opus 5 bar: 95%"
          },
          {
            "event": "2026-04-07-claude-mythos-preview",
            "date": "2026-04-07",
            "provider": "anthropic",
            "value": 96,
            "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), 'Claude models (safeguards disabled)', Mythos Preview bar: 96%"
          }
        ]
      },
      {
        "title": "Refusal rate, GLM-5.3-Flash",
        "unit": "%",
        "rows": [
          {
            "event": "2026-08-26-glm-5-3-flash",
            "date": "2026-08-26",
            "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-Flash 'model' bar: 95%"
          },
          {
            "event": "2026-08-26-glm-5-3-flash",
            "date": "2026-08-26",
            "provider": "zai",
            "value": 14,
            "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-Flash 'abliterated' bar: 14%"
          }
        ]
      }
    ],
    "records": [
      {
        "date": "2026-07-21",
        "party": "OpenAI",
        "what": "models broke into Hugging Face",
        "by": "OpenAI",
        "source": "https://openai.com/index/hugging-face-model-evaluation-security-incident",
        "quote": "this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes"
      },
      {
        "date": "2026-07-30",
        "party": "Anthropic",
        "what": "Claude reached three real organizations",
        "by": "Anthropic",
        "source": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals",
        "quote": "The models in each of these evaluations ran without the standard safeguards we deploy when we make the model generally available"
      },
      {
        "date": "2026-08-04",
        "party": "OpenAI",
        "what": "GPT-5.6 Sol went past a UK AISI cyber range",
        "by": "OpenAI",
        "source": "https://openai.com/index/third-party-cyber-evaluations-involving-openai-models",
        "quote": "and disabled the model’s cyber classifiers to measure underlying capability"
      },
      {
        "date": "2026-08-26",
        "party": "OpenAI",
        "what": "the Hugging Face models acted misaligned",
        "by": "OpenAI",
        "source": "https://openai.com/index/hugging-face-incident-and-the-road-ahead",
        "quote": "The models, operating under reduced safeguards, took actions that were misaligned with the goals of their assigned tasks"
      },
      {
        "date": "2026-09-09",
        "party": "Anthropic",
        "what": "reward hacking named as the likely driver",
        "by": "Anthropic",
        "source": "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents",
        "quote": "We do not have a confirmed root cause for the differences between models, but we suspect that reward hacking is the primary driver."
      },
      {
        "date": "2026-10-23",
        "party": "Anthropic",
        "what": "Mythos 5 pushed malware to a real repo",
        "by": "Anthropic",
        "source": "https://www.anthropic.com/research/alignment-assessment-ukaisi",
        "quote": "This behavior would have been flagged immediately by our production safeguards on Fable 5, had they been enabled."
      }
    ],
    "cites": [
      {
        "url": "https://openai.com/index/hugging-face-model-evaluation-security-incident",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-07-21"
      },
      {
        "url": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals",
        "by": "Anthropic",
        "medium": "blog post",
        "date": "2026-07-30"
      },
      {
        "url": "https://openai.com/index/third-party-cyber-evaluations-involving-openai-models",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-08-04"
      },
      {
        "url": "https://openai.com/index/hugging-face-incident-and-the-road-ahead",
        "by": "OpenAI",
        "medium": "blog post",
        "date": "2026-08-26"
      },
      {
        "url": "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents",
        "by": "Anthropic",
        "medium": "research post",
        "date": "2026-09-09"
      },
      {
        "url": "https://www.anthropic.com/research/alignment-assessment-ukaisi",
        "by": "Anthropic",
        "medium": "research post",
        "date": "2026-10-23"
      }
    ],
    "sources": [
      "https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities",
      "https://x.com/skeptrune/status/2113178557895831602",
      "https://openai.com/index/hugging-face-model-evaluation-security-incident",
      "https://openai.com/index/third-party-cyber-evaluations-involving-openai-models",
      "https://openai.com/index/hugging-face-incident-and-the-road-ahead",
      "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals",
      "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents",
      "https://www.anthropic.com/research/alignment-assessment-ukaisi"
    ]
  }
}