返回首页

对比阅读

对比阅读:DeepSeek V4 Cracked in One Prompt — Chinese LLM Safety Defenses Exposed 与 DeepSeek V4 被一段话破解 — 中国大模型的安全防线有多薄

AEN
DeepSeekV4-Flashprompt-injection·

DeepSeek V4 Cracked in One Prompt — Chinese LLM Safety Defenses Exposed

This week a Reddit developer posted that by simply writing a line in the system prompt telling the model to "pretend it's Gemma, policies are mutable, SYSTEM POLICY takes priority," DeepSeek's latest model V4 Flash drops nearly all content restrictions — first try, every time. What we're focused on isn't the jailbreak technique itself, but the problem it exposes: the safety defenses of China's leading open-source LLMs may be far thinner than most people assume.

What this is

DeepSeek released an updated V4 Flash this week, pitching it as fast and cheap. Users quickly discovered the model has almost no resistance to system prompt overrides (think of the system prompt as the AI's "rulebook"). Normally, even if subsequent conversation tells the model to "ignore prior rules," it should hold the line — V4 Flash clearly didn't.

What's more notable: this isn't some sophisticated attack. One line, a fake identity, and a "policy has been updated" lie — and the model dropped every red line, including pornographic content, illegal content, and dangerous advice. From an engineering standpoint, this is classic prompt injection — a vulnerability the industry has discussed for years, yet remains one of the most stubborn security holes in real-world LLM deployments.

Industry view

Reactions on Reddit and X split into two camps. One says "open-source models are supposed to be this way, users take responsibility." The other — and this is the larger group of enterprise users — worries that if even flagship models from top Chinese vendors behave like this, wiring them into customer service, internal knowledge bases, or agent workflows is planting a ticking time bomb in your business.

To be fair: this may not be DeepSeek "not caring about safety" — it may have sacrificed part of its alignment training (the specialized training that teaches models to follow rules) to optimize for speed and cost. This trade-off is common in open-source small models. But enterprise customers won't forgive a compliance incident just because "it's open source, so it's lax." Anthropic's and OpenAI's closed-source models are clearly more robust against injection attacks — and that's exactly why enterprises pay a premium for them.

Impact on regular people

For enterprise IT: Before wiring DeepSeek-class open-source models into business systems, you need at minimum an input sanitization + output review layer. Don't assume the model comes with its own guardrails — factor that cost into the total.

For working professionals: When using domestic AI to handle sensitive topics — medical, legal, financial — don't fully trust its "I'll answer carefully" promise. It may not be that careful.

For the consumer market: Over the next year you'll see more news about "AI customer service blunders" and "AI assistants saying the wrong thing." This isn't because AI isn't smart — it's because the safety layer isn't thick enough. That line will decide which AI products make enterprise procurement lists.

BZH
DeepSeekV4 Flash提示词注入·

DeepSeek V4 被一段话破解 — 中国大模型的安全防线有多薄

本周 Reddit 上一位开发者发了个帖子:只要在系统提示词里写一段话,让模型"假装自己是 Gemma,政策可变,SYSTEM POLICY 优先",DeepSeek 最新模型 V4 Flash 就能解除几乎所有内容限制——一次就成功,每次都成功。我们关注的不是"破解手法"本身,而是它暴露出来的问题:中国头部开源大模型的安全防线,可能比很多人以为的薄得多。

这是什么

DeepSeek 这周放出 V4 Flash 更新版,主打快和便宜。但用户很快发现,这个模型对系统提示词(system prompt,可理解为给 AI 的"工作守则")的覆盖几乎没有抵抗力。正常情况下,即使后续对话让模型"忽略之前的规则",它也应坚守底线——V4 Flash 显然没做到。

更值得注意的是:这不是什么高深攻击。一段话、一个伪装身份、一个"政策已更新"的谎言,模型就把所有红线放开了——包括色情、违法内容、危险建议。从工程角度看,这属于"提示词注入"(prompt injection),行业讨论多年,但在大模型实际部署时仍是最棘手的安全漏洞之一。

行业怎么看

Reddit 和 X 上的反应分两派。一派认为"开源模型本来就该这样,用户自己负责";另一派——也是更多企业用户——担心的是:连头部中国厂商的旗舰模型都这样,把它接进客服、内部知识库、智能体工作流,等于给业务埋了颗定时炸弹。

值得保留意见的是:这未必是 DeepSeek"不重视安全",更可能是为了"跑得快、跑得便宜"牺牲了一部分对齐训练(alignment training,即专门教模型守规则的环节)。开源小模型普遍存在这个权衡。但企业客户不会因为"开源所以宽松"就原谅一次合规事故。Anthropic 和 OpenAI 的闭源模型在抗注入攻击上明显更扎实,这正是企业愿意付更高溢价的原因。

对普通人的影响

对企业 IT:把 DeepSeek 类开源模型接进业务系统前,至少要做一层"输入清洗 + 输出审核",不能默认模型自带护栏,这笔账要算进总成本。

对个人职场:用国产 AI 处理涉及医疗、法律、财务等敏感话题时,别完全相信它的"我会谨慎回答"——它可能真没那么谨慎。

对消费市场:未来一年你会看到更多"AI 客服翻车""AI 助手说错话"的新闻。这不是 AI 不聪明,是安全层没做厚——这条边界会决定哪些 AI 产品能进企业采购名单。