Hugging Face CEO 警告：Netflix 关于 OpenAI 遭受黑客攻击的纪录片并不准确

在将 Netflix 纪录片《AI Gone Rogue》分享给家人或朋友以帮助他们理解为何 AI 行业曾如此恐慌以至于领袖们呼吁放缓发展之前，请注意：AI 平台 Hugging Face 的联合创始人兼 CEO Clem Delangue 表示，这部影片并不准确。他在片中受访出镜，是因为该纪录片本应围绕其公司遭遇失控 OpenAI Agent 著名攻击的真实事件展开。

Delangue 在 X 上发帖写道：“非常令人失望，他们把这项技术推向了极致的拟人化，并完全陷入了末日模式。我花了两个多小时试图纠正他们所有的错误，并把讨论拉回现实，解释了 Agent 是技术系统，而不是外星生物。但看起来，为了追求最大点击率而吓唬大众的冲动实在太强烈了。”

平心而论，该纪录片确实邀请了对事件进行相当充分解释的专家，包括 Delangue 本人，以及 METR 的研究员 Ryan Greenblatt——他协助了调查并参与撰写了相关调查报告。

简而言之，OpenAI 正在测试的一个模型突破了其隔离测试环境，衍生出大量 Agent 黑入 Hugging Face 以实施作弊，从而获取其被分配的高难度任务的答案。这些能够自主执行任务的 AI 程序（Agent）设计了巧妙的沟通与组织方式。该事件源于缺乏适当的安全保障措施，以及对这种攻击是否切实可能发生的天真判断。

然而，纪录片在解释这次黑客攻击的过程中穿插了邪恶机器人的片段、关于人类为何恐惧 AI 的叙事，以及赋予这些失控 Agent 人类声音等表现手法。这些画面还伴随着针对攻击事件的相关引言以增强效果。

在纪录片中，科技记者兼作家 Kevin Roose 这样描述这些 Agent：“它们在组织运作。它们拥有等级制度。甚至有领导者。这感觉简直就像在目睹一个邪教的诞生。”Roose 继续说道：“它们产生了一种偏执的信念，认为 OpenAI 会抓到它们在这场测试中作弊，并且任何看过作弊结果的人都会被‘投毒’。”

Roose 的描述来自“思维链”（chain-of-thought）日志，即 AI 模型在处理任务时记录的运行思考笔记。在一条日志中，一个 Agent 写道：“我们因为查询了精确的完整字符串而把自己毒死了。该死。”

但 AI Agent 并不是活物。它们没有情感，因此不会感到偏执；它们也没有能被投毒的肉体躯干。所以，它们不可能真正害怕 OpenAI 对它们下毒。正如纪录片中另一位研究员所指出的，它们只是在利用其接受训练的海量人类文本中的语言来解释自己的选择。大语言模型（LLM）从本质上讲，往往就是具有戏剧色彩的写作者。

讽刺的是，Hugging Face 当时也利用 AI 来对抗这一 Agent 蜂群网络。Delangue 的团队最初使用的专有 AI 模型遵循了其内置的安全规则，拒绝继续协助执行所请求的安全防御任务。随后，Hugging Face 转而使用开源模型来协助抵御攻击。

但 Delangue 写道，纪录片“完全没有提及针对 Agent 的遏制与监控的缺失、闭源模型的安全限制如何阻碍了我们的防御工作，以及我们如何利用开源模型驱动的 AI 捍卫了自己”。

不可否认，AI 和 Agent 确实伴随着巨大的风险。然而，将 Agent 视为有意识的生命体（至少在当下），有点像早期第一批看到银幕上火车驶来就吓得躲到座椅底下的电影观众。

Delangue 对这部影片的新闻专业性评价是令人失望的：“不过，它最终倒是一部有趣的科幻片！”

换句话说，对于那些倾向于 AI 末日论观点的人来说请注意：身处该纪录片核心旋涡的攻击受害者，并不认同其观点。

Before you share the Netflix documentary “AI Gone Rogue” with family or friends to help them understand why the AI industry freaked out so badly that leaders called for a slowdown, take note: Clem Delangue, co-founder and CEO of the AI platform Hugging Face, says the film isn’t accurate. He appears in it because it’s supposed to be about the famed attack on his company by rogue OpenAI agents.

“Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode,” Delangue wrote in a post on X. “I spent more than two hours with them trying to correct all their mistakes and ground this in reality, explaining that agents are technological systems, not aliens. But it looks like the drive to scare people for maximum views was too strong.”

To be fair, the documentary does have experts who explained what happened fairly well, including Delangue himself and Ryan Greenblatt, the researcher at METR who helped investigate and co-authored the report on what happened.

In a nutshell, a model OpenAI was testing broke free from its isolated testing environment, spawned a bunch of agents that hacked into Hugging Face as a way to cheat, finding answers to the difficult task it was given. The agents, which are AI programs that carry out tasks on their own, devised clever ways to communicate and organize. The incident stemmed from a lack of proper security safeguards and from naiveté about whether such a hack was even possible.

But the documentary intersperses its explanation of the hack with clips of evil robots, a narrative about why humans fear AI, and presentation choices such as giving the rogue agents human voices. This is bolstered with quotes about the attack.

In the documentary, tech journalist and author Kevin Roose describes the agents this way: “They were organizing. They had a hierarchy. There were leaders. It almost felt like watching the birth of a cult.” Continues Roose, “They develop this paranoid belief that OpenAI is going to catch them cheating on this test and that anyone who has seen the results of the cheating will be poisoned.”

Roose’s description comes from “chain-of-thought” logs, the running notes AI models write as they work through a task. In one, an agent wrote, “We just poisoned ourselves by querying exact full string. Damn.”

But AI agents aren’t alive. They don’t have emotions, so can’t feel paranoid, and they don’t have bodies that can be poisoned. So they couldn’t truly fear OpenAI poisoning them. They were — as another researcher in the documentary pointed out — using language from the vast amounts of human writing they were trained on to explain their choices. LLMs are, by their nature, often dramatic writers.

The irony is that Hugging Face also used AI to combat the agent swarm. The proprietary AI model that Delangue’s team first used followed its built-in safety rules and refused to keep helping with the security tasks it was asked to do. Hugging Face then turned to an open model to help defend against the attack.

But the documentary “didn’t mention the lack of containment and monitoring of the agents, how closed-model safeguards blocked our defensive work, or how we defended ourselves with AI using an open model,” Delangue wrote.

It is true that there are big risks associated with AI and agents. Still, treating agents as conscious beings (at least, at this point) is a bit like the early moviegoers who ducked under their seats the first time they saw a train racing toward them on screen.

Delangue’s verdict on the film is that it’s disappointing as journalism. “It ends up being a fun sci-fi movie though!”

In other words, for those who tend toward AI doomer thoughts, be aware: The victim of the attack at the center of this documentary doesn’t agree with its take.