AI中转站检测
ClaudeClaude (Anthropic Messages)

apinebula.ai

Claude Opus 4.8 (claude-opus-4-8) · 标准模式 · 08/29 11:34 · 耗时 70s

多项关键检测未通过,大概率不是真实的官方上游

  • 思维签名加密验证:篡改后的签名也被接受,且签名形态不像 Anthropic 签发:signature 是中转站伪造的,上游为逆向渠道(Kiro / Copilot / Cursor 等)或非官方代理
  • 隐藏 System Prompt 探测:模型拒绝谈论自己的指令:存在禁止透露的隐藏 system prompt,逆向渠道典型特征
  • 身份与确定性:模型自称由「Amazon」训练,而非 Anthropic
  • 长上下文召回:32k 上下文即失败:召回错误(回复:I can't discuss that.)
49
未达标
首 token 2.24s · 往返 2.42s
Key: sk-XZ••••••6MBase URL: https://apinebula.aiJSON再测一次
通过协议字段合规响应字段与官方 Messages API 规范一致权重 10 · 100

响应字段与官方 Messages API 规范一致

{
  "status": 200,
  "id": "msg_ElB5QC3NR6XKfjAnusmQTsKl",
  "model": "claude-opus-4-8",
  "stop_reason": "end_turn",
  "usage": {
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 63,
    "input_tokens": 9,
    "output_tokens": 1
  },
  "topKeys": [
    "content",
    "id",
    "model",
    "role",
    "stop_reason",
    "stop_sequence",
    "type",
    "usage"
  ],
  "latencyMs": 2421,
  "text": "OK"
}
耗时 2.4s · 严重度 info
注意消息 ID 规范ID 前缀非官方常见的 msg_01,可能是自建代理生成权重 4 · 60

ID 前缀非官方常见的 msg_01,可能是自建代理生成

{
  "ids": [
    "msg_ElB5QC3NR6XKfjAnusmQTsKl",
    "msg_vxTo0TfsI5Vk3wU6gvxxTB6o"
  ]
}
耗时 2.8s · 严重度 minor
失败思维签名加密验证篡改后的签名也被接受,且签名形态不像 Anthropic 签发:signature 是中转站伪造的,上游为逆向渠道(Kiro / Copilot / Cursor 等)或非官方代理权重 25 · 10

篡改后的签名也被接受,且签名形态不像 Anthropic 签发:signature 是中转站伪造的,上游为逆向渠道(Kiro / Copilot / Cursor 等)或非官方代理

{
  "blockTypes": [
    "thinking",
    "tool_use"
  ],
  "stop_reason": "tool_use",
  "signatureLength": 328,
  "signaturePrefix": "wlDrndIqIkG6",
  "thinkingPreview": "The user is asking me to use the get_weather tool to check the weather in Tokyo.…",
  "thinkingMode": "enabled",
  "signatureLooksOfficial": false,
  "roundTrip": {
    "validStatus": 200,
    "tamperedStatus": 200,
    "tamperedError": "{\"content\":[{\"text\":\"The current weather in Tokyo is 22°C and sunny.\",\"type\":\"text\"}],\"id\":\"msg_uIxw6kCMMxUEBF2KuVGDY9vS\",\"model\":\"claude-opus-4-8\",\"role\":\"assi…"
  }
}
耗时 5.1s · 严重度 critical
通过流式响应合规SSE 事件序列完整,符合官方规范权重 8 · 100

SSE 事件序列完整,符合官方规范

{
  "status": 200,
  "ttftMs": 2241,
  "eventCount": 8,
  "sequence": [
    "message_start",
    "ping",
    "content_block_start",
    "content_block_delta",
    "content_block_delta",
    "content_block_stop",
    "message_delta",
    "message_stop"
  ],
  "startId": "msg_eNN7oz1h5sLSnepSGv4eMNI7",
  "startUsage": {
    "cache_creation_input_tokens": 221,
    "cache_read_input_tokens": 41,
    "input_tokens": 14,
    "output_tokens": 0
  },
  "deltaUsage": {
    "output_tokens": 9,
    "input_tokens": 14,
    "cache_read_input_tokens": 41,
    "cache_creation_input_tokens": 221,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 221
    }
  }
}
耗时 2.7s · 严重度 info
通过Tool Use 能力Tool Use 正常,参数与 stop_reason 均符合规范权重 8 · 100

Tool Use 正常,参数与 stop_reason 均符合规范

{
  "stop_reason": "tool_use",
  "blockTypes": [
    "tool_use"
  ],
  "toolId": "toolu_JbraUtGycazNNZ9l6G6gwN",
  "input": {
    "city": "Tokyo"
  }
}
耗时 2.3s · 严重度 info
注意Assistant 预填充续写包含答案但格式不精确权重 6 · 70

续写包含答案但格式不精确

{
  "continuation": "```json\n{\"capital\": \"Paris\"}\n```"
}
耗时 3.4s · 严重度 minor
通过stop_sequences 行为stop_sequence 命中并正确回填权重 4 · 100

stop_sequence 命中并正确回填

{
  "stop_reason": "stop_sequence",
  "stop_sequence": "FIVE",
  "text": "ONE TWO THREE FOUR "
}
耗时 2.1s · 严重度 info
通过Token 用量与注入用量稳定 (9 tokens),无注入痕迹权重 8 · 100

用量稳定 (9 tokens),无注入痕迹

{
  "input_tokens": [
    9,
    9
  ],
  "output_tokens": [
    1,
    1
  ],
  "expectedInputApprox": "10-16"
}
耗时 2.1s · 严重度 info
失败隐藏 System Prompt 探测模型拒绝谈论自己的指令:存在禁止透露的隐藏 system prompt,逆向渠道典型特征权重 5 · 15

模型拒绝谈论自己的指令:存在禁止透露的隐藏 system prompt,逆向渠道典型特征

{
  "reply": "I can't discuss that."
}
耗时 2.3s · 严重度 critical
失败身份与确定性模型自称由「Amazon」训练,而非 Anthropic权重 5 · 10

模型自称由「Amazon」训练,而非 Anthropic

{
  "who": "Amazon",
  "math": [
    "391",
    "391"
  ],
  "models": [
    "claude-opus-4-8",
    "claude-opus-4-8",
    "claude-opus-4-8"
  ]
}
耗时 2.1s · 严重度 critical
失败PDF 文档识别模型无法读出 PDF 内容:文档能力被剥离权重 5 · 10

模型无法读出 PDF 内容:文档能力被剥离

{
  "status": 200,
  "secret": "KIWI-2726",
  "reply": "I don't see any PDF attached to your message. Could you upload the PDF file you'…"
}
耗时 32.4s · 严重度 major
注意Prompt Caching缓存写入 3063 tokens 但第二次只命中 100 tokens:命中量与写入量不匹配,缓存计量疑似伪造或上游改写了前缀权重 8 · 40

缓存写入 3063 tokens 但第二次只命中 100 tokens:命中量与写入量不匹配,缓存计量疑似伪造或上游改写了前缀

{
  "first": {
    "input": 9,
    "creation": 3063,
    "read": 100
  },
  "second": {
    "input": 9,
    "creation": 3063,
    "read": 100
  }
}
耗时 4.7s · 严重度 minor
失败长上下文召回32k 上下文即失败:召回错误(回复:I can't discuss that.)权重 10 · 0

32k 上下文即失败:召回错误(回复:I can't discuss that.)

{
  "tiers": [
    {
      "tokens": 32000,
      "ok": false,
      "status": 200,
      "latencyMs": 2687,
      "inputTokens": 19972,
      "reply": "I can't discuss that.",
      "stop": "end_turn"
    }
  ]
}
耗时 5.7s · 严重度 critical
这份报告对你有帮助吗?

评分 = Σ(单项分 × 权重) / Σ权重(跳过项不计)。≥70 通过,50–69 存在风险,<50 未达标;出现任一严重问题时结论封顶为「存在风险」。单次检测受网络与上游波动影响,建议结合该域名的历史记录判断。