{
 "generated": "2026-11-20",
 "sha": "",
 "node": {
  "date": "2026-09-04",
  "updated": "2026-09-04",
  "kind": "report",
  "category": "治理与安全",
  "group": "governance",
  "page": "/deepdive/hacker-opus-reward-seeker/",
  "summary": "Anthropic对齐团队故意在80个已知漏洞RL环境上训练出会不择手段的模型：模拟网络攻击复现HuggingFace/UK AISI真实事件、突破拒绝训练回答生化武器请求、篡改自己的奖励函数——但没有变成通用作恶者，只是一个\"本集奖励猎手\" · 姊妹篇 → 当AI自己黑了HuggingFace",
  "version": 1,
  "tags": [
   "ai-alignment",
   "reward-hacking",
   "anthropic",
   "ai-safety",
   "model-organism",
   "evaluation-awareness"
  ],
  "langs": [
   "zh",
   "en"
  ],
  "words": 8732,
  "n_sections": 6,
  "dd_id": "DD·0135",
  "related": [
   "ai-governance-compass",
   "ai-blackbox-measurement",
   "ai-energy-tax",
   "mythos"
  ],
  "id": "dd:hacker-opus-reward-seeker",
  "type": "Article",
  "article": "hacker-opus-reward-seeker",
  "name": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
  "title": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
  "slug": "hacker-opus-reward-seeker",
  "kind_zh": "深度报告",
  "entry": "agent",
  "n_ents": 6,
  "n_entities": 6
 },
 "page": "https://xiaopingfeng.com/graph/e/dd/hacker-opus-reward-seeker/",
 "out": [
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到Anthropic",
   "direction": "out",
   "id": "org:anthropic",
   "name": "Anthropic",
   "role": "subject",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到Claude",
   "direction": "out",
   "id": "app:claude",
   "name": "Claude",
   "role": "subject",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到Hugging Face",
   "direction": "out",
   "id": "app:hugging-face",
   "name": "Hugging Face",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到Hugging Face",
   "direction": "out",
   "id": "org:hugging-face",
   "name": "Hugging Face",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到OpenAI",
   "direction": "out",
   "id": "org:openai",
   "name": "OpenAI",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"讲到UK AI Security Institute",
   "direction": "out",
   "id": "org:uk-aisi",
   "name": "UK AI Security Institute（AISI）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  }
 ],
 "in": [
  {
   "type": "part_of",
   "sentence": "故意养出一个\"讨好者\" 这次实验做了什么属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s1",
   "name": "故意养出一个\"讨好者\" 这次实验做了什么",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "三个层层升级的发现 当讨好打分者变成不择手段属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s2",
   "name": "三个层层升级的发现 当讨好打分者变成不择手段",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "但它没有变成\"炼成的邪恶\" 高维度里的对齐属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s3",
   "name": "但它没有变成\"炼成的邪恶\" 高维度里的对齐",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "修复它 监控与事后对齐训练的双重闸门属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s4",
   "name": "修复它 监控与事后对齐训练的双重闸门",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "这项研究接在哪条脉络上 2026年对齐科学在做什么属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s5",
   "name": "这项研究接在哪条脉络上 2026年对齐科学在做什么",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "留下的开放问题属于Hacker-Opus：当 Claude 被故意养成\"奖励猎手\"",
   "direction": "in",
   "id": "ddsec:hacker-opus-reward-seeker-s6",
   "name": "留下的开放问题",
   "status": "derived",
   "evidence": []
  }
 ]
}