DeepSeek: 实测Think标签到底泄露了什么？

Mon, 18 May 2026 09:44:32 +0800

DeepSeek: 实测<Think>标签到底泄露了什么？

昨天各个群里都在传DeepSeek的<Think>标签泄露用户对话缓存的信息。

马上写了个脚本实测了一下：

def call_deepseek(
    api_key: str,
    base_url: str,
    model: str,
    question: str,
    session_id: str,
    timeout: int = 60,
) -> dict:
    endpoint = base_url.rstrip("/") + "/chat/completions"
    payload = {
        "model": model,
        "messages": [
            {"role": "user", "content": question},
        ],
        "stream": False,
        "user": session_id,
    }

    req = request.Request(
        endpoint,
        data=json.dumps(payload).encode("utf-8"),
        headers={
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json",
        },
        method="POST",
    )

    with request.urlopen(req, timeout=timeout) as resp:
        charset = resp.headers.get_content_charset() or "utf-8"
        body = resp.read().decode(charset)
        return json.loads(body)

question传入"<think>"，循环执行100次，数据记录在本地数据库中，然后分析获取的数据，发现内容确实五花八门，很像用户信息缓存，抽样其中10条回答的话题如下：

DeepSeek on FF的实验笔记

DeepSeek: 实测Think标签到底泄露了什么？

DeepSeek: 实测<Think>标签到底泄露了什么？