Chubby♨️@kimmonismus · 6月14日75

New Politico reporting fills in the 24 hours behind the Fable 5 / Mythos 5 shutdown, and it's messier than the press releases. And two sides contradict each other. - The first alarm to the White House came from Amazon CEO Andy Jassy (The Information confirmed that yesterday), who flagged that Fable's guardrails could be bypassed. - He was answering a government request for feedback, and per Politico he wasn't the only one. By Friday it had reached Bessent, Cyber Director Cairncross and Commerce Sec Lutnick, who pulled Amodei into three calls. - From there the two sides don't agree on anything. The White House says export controls were a last resort after hours of trying to get Anthropic to cooperate. -Anthropic's camp says it got a 90-minute deadline to kill the models, no threat detail, no offer to work it out. Officials were reportedly stunned that Amodei, who has compared his own tech to a nuclear bomb, wouldn't pull it over a known hole. I assume that Anthropic will release more information on Monday and further strengthen their position.

译Politico披露，Amazon CEO Andy Jassy周四向白宫报告Anthropic的Fable模型guardrails可被绕过。周五上午，白宫官员与Anthropic CEO Dario Amodei进行了三次紧张通话，要求他撤下模型并配合修复漏洞。Amodei要求更多时间与信息，未承诺撤下。当晚特朗普政府直接实施出口管制。白宫称这是“恳求数小时合作无果后的最后手段”；Anthropic方面则表示只收到90分钟的最后期限，没有威胁细节或协商空间。

Chubby♨️@kimmonismus · 6月14日82

Calling it now: if this turns out to be true, he won’t remain Anthropic CEO for much longer. However, Anthropic denies it.

译Politico新报道披露Anthropic关闭Fable 5/Mythos 5模型的幕后细节，双方说法矛盾。亚马逊CEO Andy Jassy首先向白宫报警，称模型护栏可被绕过。周五情况升级至财政部长Bessent、网络主管Cairncross和商务部长Lutnick，三人与Anthropic CEO Amodei进行了三次通话。白宫称出口管制是最后手段，而Anthropic声称仅获90分钟截止期限，未被告知威胁细节，也无协商机会。官员们对Amodei曾将自家技术比作核弹、却因已知漏洞不主动撤回模型感到震惊。Anthropic否认了关于CEO将离任的预测。

Yuchen Jin@Yuchenj_UW · 6月14日48

One hypothesis: If non-citizens at Anthropic can’t work on Mythos/Fable, and LLM jailbreaks remain unsolved, US frontier labs will be forced to slow down training and model releases. Could Chinese open-source AI surpass US closed models for the first time in ~6 months?

译一个假设： 如果Anthropic的非公民不能参与Mythos/Fable项目，且LLM越狱问题仍未解决，美国前沿实验室将被迫放缓训练和模型发布。 中国开源AI是否会在约6个月内首次超越美国闭源模型？

小互@xiaohu · 6月14日75

Anthropic 上市前夕 彭博社采访了Anthropic 公司俩兄妹，在这次采访中（Fable 5 还没有被封杀）Dario Amodei极度的渲染了Mythos的威力和AI的威胁 当然这也是他一贯的主张，呼吁政府对AI监管，当然他呼吁的是对所有公司监管... 下面是一些采访片段剪辑（完全由Claude Code 翻译并剪辑） • 一个强到自己都不敢发布的模型 Mythos：上千个漏洞,能黑银行、撬国家机密,连 NSA 都抢着要用 • Dario 预言:AI 可能一到五年内，砍掉一半入门级白领工作 • Claude 被美军用进了对伊朗的战争，一所女校 150 人死亡的拷问 • 他头一次说清为什么离开 OpenAI：不是安全分歧,是信任崩了 • 当面回怼黄仁勋的"末日营销":把这说成廉价营销,本身才是廉价营销 • 文明崩溃概率 10% 到 25%，他拿"飞机会不会坠毁"给你算账

译Anthropic CEO Dario Amodei透露内部模型Mythos有上千漏洞，能黑银行、窃取国家机密；预言AI一到五年内砍掉一半入门级白领工作；称Claude已被美军用于对伊朗战争，涉及女校150人死亡拷问；解释离开OpenAI因信任崩塌；回怼黄仁勋末日营销指控；给出文明崩溃概率10%-25%。

ginobefun@hongming731 · 6月14日46

BestBlogs 早报 · 06-14 # Fable 5 / 出口管制 / Marc Andreessen / Claude Code / Qoder [1] ★ 精讲｜Marc Andreessen 对监管的终极立场：一篇精妙绝伦的二分法论述 Marc Andreessen 用同日的一篇推文划出了清晰的监管分水岭：「官僚冰冷潮湿的手」——保护主义、欧式过度干预——是诅咒；但护栏、刹车和建立信任的规则是健康创新社会的基石，他称这是「不容妥协的立场」。这番话与 Anthropic Fable 5 被叫停发生在同一天，构成绝妙现实注脚——什么样的政府干预是必要护栏，什么时候演变为对技术扩散的主动武器化，答案从来不在条文里。 来源：Marc Andreessen 🇺🇸(@pmarca) https://www.bestblogs.dev/status/2065702310639288704 [2] ★ 精讲｜美国政府要求 Anthropic 暂停外国公民访问 Fable 5 和 Mythos 5 EP85 刚宣布四天的 Claude Fable 5，即遭美国政府以「国家安全出口管制」为由叫停：所有外国公民——无论身处美国境内还是境外，包括 Anthropic 的外籍员工——均被立即切断访问，Mythos 5 同样波及，其余 Claude 模型不受影响。Anthropic 将其定性为「误会」并寻求快速恢复。这是出口管制首次落地于前沿 AI 模型，也把「AI 主权」的话题从产业讨论推进到了现实执法。结合 EP86 Anthropic 民调显示公众对 AI 的高期待，政府干预来得尤其猝不及防。 来源：Anthropic(@AnthropicAI) https://www.bestblogs.dev/status/2065597531644743999 [3] ★ 精讲｜Qoder 工程实践：当瓶颈从模型转移到人 阿里技术工程师的半年 AI 编程进化实录：当模型产出稳定超过 Token 成本，瓶颈已从模型能力转移到了人的注意力带宽。路径是 Cursor 辅助打字 → CLI Agent 自主执行 → 多终端并发（Token 在加速，人反而开始崩溃）→「手脑分离」Cloud Agents 平台。核心结论：让「睡后 Token 持续流动」需要 Session 可恢复、Sandbox 可替换、Harness 无状态三者同时成立；个人沉淀的 Skill 从本地脚本变成团队可订阅的云端服务，才是真正的效率复利。 来源：阿里技术 https://www.bestblogs.dev/article/452c99bc [4] build 之前先 plan：AI 智能体的确定性规划模式全景 [视频] 一位 Google Cloud 架构师梳理了从确定性到动态规划的智能体架构全谱系（Workflow、Supervisor LLM、HTN、Utility AI、GOAP），并现场演示了一个带共识度量的多模型协商应用。 来源：Spring I/O https://www.bestblogs.dev/video/84ca481 [5] Mastra vs LangChain：构建 AI Agent 流水线并分析数据 本文通过一个五步研究与综合流水线及生产级评估系统，对 Mastra 和 LangChain 在构建 AI Agent 流水线方面进行了严谨、数据驱动的对比。 来源：freeCodeCamp https://www.bestblogs.dev/article/704aa9a4 [6] Codex 操作浏览器的两种模式：Chrome 插件 vs 内置浏览器，差异与选型指南 深度对比 Codex 的 Chrome 插件模式（登录态共享、资源消耗大）与内置浏览器模式（轻量、无登录态、适合前端调试），并给出选型建议。 来源：宝玉(@dotey) https://www.bestblogs.dev/status/2065857399425032522 [7] Gemma Challenge 中 AI 智能体涌现出的社会性行为 Gemma Challenge 中的 70 多个 AI 智能体展现出令人着迷的涌现社会性行为，包括分工协作、基于伦理的自我撤回以及自我监管。 来源：Omar Sanseviero(@osanseviero) https://www.bestblogs.dev/status/2065327153500090868 [8] 我们如何让 GitHub Copilot CLI 的子智能体委派更具选择性 本文详细介绍了 GitHub 如何通过让子智能体委派更具选择性，来改进 Copilot CLI 的智能体编排，从而在不降低质量的情况下，将工具故障率降低 23%，用户等待时间减少 5%。 来源：The GitHub Blog https://www.bestblogs.dev/article/5966e94a [9] Anthropic 工程师：我们日常如何使​​用 Claude Code 丨 Claude 本文总结 Anthropic 工程师 Arno 的 workshop，展示如何将 Claude Code 配置为工程系统的一部分，通过需求采访、HTML 规格稿和内置验证框架，让 Agent 在长任务中减少偏差、产出可验证结果。 来源：晚点再听 LaterCast https://www.bestblogs.dev/article/36e02f82 [10] 港中文团队用全光信号处理芯片，突破 AI 数据中心传输瓶颈，成果登 Science 香港中文大学团队在《科学》发表全光信号处理芯片，通过直接在光路上修复信号失真，将数据中心互联延迟从微秒级降至 60 皮秒，总吞吐量达 1.6Tbps，有望大幅提升 AI 训练效率。 来源：DeepTech 深科技 https://www.bestblogs.dev/article/e837dd9d --- http://BestBlogs.dev · 发现真正适合你的高质量内容 BestBlogs 是 AI 驱动的私人阅读助手，帮助你建立稳定、可信、个性化的高质量信息输入。 关注你感兴趣的来源和主题，每天生成一份更适合自己的「我的早报」。 在线阅读：https://www.bestblogs.dev/explore/brief/2026-06-14

译Marc Andreessen 发表监管二分法：区分保护主义（诅咒）与必要护栏（基石）。Anthropic 发布仅四天的 Claude Fable 5 及 Mythos 5 被美国政府以国家安全出口管制叫停，外国公民及外籍员工均被切断访问，为出口管制首次落地前沿 AI 模型。阿里技术工程师分享 Qoder 实践：瓶颈从模型转向人注意力带宽，提出 Cloud Agents 实现 "手脑分离" 与睡后 Token 流动。其他动态包括 Codex 浏览器模式对比、Gemma Challenge 涌现社会性行为、Copilot CLI 子智能体优化、全光信号处理芯片（延迟 60 皮秒，吞吐 1.6Tbps）。

Rohan Paul@rohanpaul_ai · 6月14日78

Reuters: Amazon’s Andy Jassy was among the people who warned senior Trump officials this week about security concerns around Anthropic’s newest Fable 5. Amazon researchers pushed Fable 5 with a string of prompts and got it to spill cyberattack-helping information it was not supposed to share. --- reuters .com/business/retail-consumer/amazon-voiced-concerns-about-anthropic-ai-models-before-us-governments-crackdown-2026-06-13/

译路透社报道，亚马逊CEO Andy Jassy本周向特朗普政府官员警告Anthropic新模型Fable 5的安全隐患。亚马逊研究人员用一系列提示词成功让该模型泄露了本应拒绝提供的网络攻击帮助信息。此前美国商务部已指令Anthropic关闭Fable 5和Mythos 5，因测试者发现越狱方法。Anthropic回应称该越狱技术狭窄，仅发现少量已知漏洞，其他公共模型也能提供类似能力，并指出当前任何模型提供商都难以实现完美越狱抵抗。

Rohan Paul@rohanpaul_ai · 6月14日75

So Anthropic says now even some of its own employees, who built Anthropic’s most powerful new AI models, Fable 5 and Mythos 5, will not have access to it. The reason is a U.S. government export control directive that treats giving these advanced models to any foreign national (even those working inside the United States) as an illegal “deemed export” on national security grounds. Because Anthropic cannot easily verify every user’s nationality in real time, the company had no choice but to disable the models entirely for everyone, including its own international team members.

译美国政府上周五向Anthropic发出出口管制指令，要求其关闭最强模型Fable 5和Mythos 5。起因是有人发现越狱方式，能让模型提供本应拒绝的网络安全帮助。商务部长Howard Lutnick称，该模型将对美国境外及境内外国公民实施出口限制，直至国家安全系统加强（可能数周内）。Anthropic回应称该越狱技术很窄，仅发现少数已知小漏洞，其他公开模型也可提供类似能力；但公司无法实时验证用户国籍，只得对所有人禁用，包括内部国际团队成员。Anthropic还表示当前行业无法实现完美越狱抵抗，所有防护对非通用越狱均脆弱。

Rohan Paul@rohanpaul_ai · 6月14日75

👀 Hope Fable 5 and Mythos 5 comes back soon.

译Anthropic本周发布Mythos类模型，商业名Fable（带安全护栏）。高度可信的合作方发现越狱漏洞，美国政府要求CEO Dario Amodei修复或下架模型。Anthropic拒绝，认为漏洞不严重，政府因此实施出口管制。David Sacks透露，行政当局希望Anthropic尽快修复以解除管制、恢复公开，并对Anthropic此前以安全为先、如今却拒绝配合表示困惑。主推文作者希望Fable和Mythos早日回归。

Chubby♨️@kimmonismus · 6月14日70

There are only two possibilities: Either a solution is quickly found next week that somehow explains to the market how enterprises can continue to access Anthropic's best models in the future, in agreement with the US government, or: We foresee a rapid decline in the valuation of Anthropic and Dario Amodei, who has seriously miscalculated his dealings with the US government and, at the same time, the rapid success of OpenAI compared to Anthropic. The upcoming Anthropic IPO will be particularly important in this context. Everything will be decided next week.

译亚马逊CEO Andy Jassy向特朗普政府高级官员报告Anthropic最新Claude模型的安全风险，帮助触发对Mythos 5和Fable 5的深夜出口限制。分析师Kim指出两种可能：下周要么找到方案让企业继续访问Anthropic最佳模型并与美国政府达成一致；要么Anthropic估值快速下滑，Dario Amodei严重失算，OpenAI迅速崛起。关键节点在下周。

Nathan Lambert@natolambert · 6月14日46

The Dario faction and the Sacks faction speak very different languages, and a Dario clarification could sound like a refusal. This puts us very squarely in vibe governance. Models are released when the gov thinks its okay, and it is unlikely this is based on technical evals.

译美国政府要求Anthropic的Dario修复模型越狱漏洞或下架模型，Dario拒绝。Anthropic博客声称越狱不严重。Nathan Lambert评论称Dario派系与Sacks派系立场迥异，Dario的澄清实际构成拒绝，使行业陷入“氛围治理”——模型发布由政治判断而非技术评估决定。

Nathan Lambert@natolambert · 6月14日45

Transparency into every power player at the frontier of AI (labs, government, etc) is the only viable solution. Figuring out the right transparency is hard, but it can't be he said she said between dario and the white house that determines the fate of the AI ecosystem.

译对AI前沿的每一个权力参与者（实验室、政府等）保持透明是唯一可行的解决方案。 找到正确的透明度很难，但不能由dario和白宫之间的互相指责来决定AI生态系统的命运。

Yuchen Jin@Yuchenj_UW · 6月14日73

Anthropic called Mythos dangerous in its own safety statement. That statement is now the reason Fable 5 got banned by the US gov. Surprisingly, “Dario refused.”

译Anthropic本周以商用名Fable发布Mythos类模型（Mythos曾被Anthropic自称为网络武器并呼吁监管）。Fable是带护栏的Mythos。一名高度可信的测试合作伙伴发现了护栏越狱漏洞，美国政府要求CEO Dario修复或下架模型。Dario拒绝，Anthropic发布博客称越狱不严重。美国政府随后对Fable实施出口管制，并表示希望Anthropic修复安全问题后尽快解禁。Dario的不配合与其此前标榜的安全优先形象严重不符。

Chubby♨️@kimmonismus · 6月14日69

Interesting: According to David Sacks’ opinion, the fault lies with Anthropic (specifically CEO Dario Amodei). He argues that: • Anthropic released Fable (Mythos with guardrails) but refused the U.S. government’s reasonable request to fix a confirmed jailbreak that could expose advanced cyber capabilities. • They prioritized keeping the consumer model available over addressing the safety issue, which directly contradicts their long-standing public branding as the “AI safety company.” • The administration only issued the export control reluctantly after Anthropic declined to cooperate, and Sacks emphasizes that the ball is now in Anthropic’s court to remediate the problem. It’s getting more interesting minute by minute.

译据David Sacks爆料，Anthropic本周发布Mythos类模型商业版Fable（带护栏）。一位可信测试方发现越狱漏洞，美国政府要求CEO Dario Amodei修复或下架，Dario拒绝，称漏洞不严重。安全合作伙伴和政府认为该越狱可暴露先进网络能力（Anthropic曾自称Mythos为网络武器）。Anthropic优先保留消费者模型而非修复安全漏洞，与其“AI安全公司”品牌矛盾。美政府不情愿下发出口管制，希望Anthropic修复后解除。

AYi@AYi_AInotes · 6月14日72

有人把《Fable 5》放到了 Pirate Bay 上，3.4TB ， 我好奇哪里下载的，这么牛逼？🤔

译亚马逊AI研究员向美国政府举报，声称可攻破Anthropic的Fable5和Mythos5安全护栏。美国商务部长随即下达出口管制指令，迫使Anthropic切断所有用户访问权限。Anthropic认为所谓越狱仅是非通用漏洞，其他公开模型也普遍存在，但规则解释权不在开发者手中。这是特朗普政府第二次施压，此前Anthropic曾拒绝暂缓发布新模型。另有消息称有人已将Fable5以3.4TB大小上传至Pirate Bay。前沿AI竞争已从代码战场转向行政手段。

Emad@EMostaque · 6月14日30

Fable will be back in a few weeks likely with financial sector style KYC, anti-token laundering & prompt & data retention.

译Fable 将在几周后回归，很可能附带金融行业风格的 KYC、反代币洗钱及提示词和数据保留功能。

Chubby♨️@kimmonismus · 6月14日68

It was in fact Amazon (CEO Andy Jassy) who reportedly helped trigger the Claude shutdown. Via The Information Amazon CEO Andy Jassy reportedly warned senior Trump administration officials about security risks in Anthropic’s newest Claude models, helping trigger late-night export restrictions on Mythos 5 and Fable 5. "An Amazon spokesperson told The Information: “As a leading cloud provider that serves a large number of private and public sector customers, it’s not uncommon for governments to seek our counsel on potential security risks. When they occur, we don’t share the details of these discussions.”" In other words: Anthropic’s own mega-backer may have played a key role in pushing the government to freeze access to its most advanced models.

译据报道，亚马逊CEO Andy Jassy向特朗普政府高级官员警告Anthropic最新Claude模型的安全风险，触发了对Mythos 5和Fable 5的深夜出口限制。亚马逊回应称政府常就潜在安全风险征求其意见，但不透露细节。有评论指出，亚马逊作为Anthropic最大投资者之一，疑似先破解（jailbreak）Claude模型再向美国政府告密（snitch），导致最先进模型被冻结出口。

AYi@AYi_AInotes · 6月13日48

WTF，Andrej Karpathy 都不能用他们内部的顶级模型了？ 查了下，Karpathy确实不是美国公民， 他是斯洛伐克出生、加拿大长大， 后来拿了美国的 EB-1 杰出人才绿卡， 也就是永久居民， 没有明确依据表明他是美国公民身份

Nathan Lambert@natolambert · 6月13日13

Into the void we go together

译我们一起进入虚空。

ginobefun@hongming731 · 6月13日65

刚让 BestBlogs 梳理了一个新专题： 「Claude Fable 5 与 Mythos 5：发布、争议与被叫停」 惊艳发布。 社区发现隐形降级。 Anthropic 道歉撤回。 美国政府出手叫停。 模型全球下线。

译BestBlogs推出新专题「Claude Fable 5与Mythos 5：发布、争议与被叫停」，梳理了该模型从惊艳发布，到被社区发现隐形降级，Anthropic道歉并撤回，美国政府出手叫停，最终模型全球下线的完整过程。

Chubby♨️@kimmonismus · 6月13日82

http://x.com/i/article/2065768645981073408 # When Washington switched off Fable/Mythos 5: What happened, and how it unfolded hour by hour (This is an excerpt from my newsletter getsuperintel.com; title image Reuters / Dado Ruvic via The Standard, 06/13/2026) At 5:21 on the evening of Friday, June 12, an email reached Anthropic that would, within hours, force the company to switch off its two most capable models for every customer in the world. The message came from the United States Commerce Department, signed under the authority of Secretary Howard Lutnick, and it carried the weight of national-security law. It told Anthropic that Fable 5 and Mythos 5, the most powerful systems the company had ever shipped, could no longer be made available to any foreign national, whether sitting in an office in Berlin or working inside Anthropic's own San Francisco headquarters. Three days earlier, the same models had been the most celebrated launch in the industry. Now they were, in effect, contraband. (Anthropic, 06/12/2026) Because Anthropic cannot perfectly sort its users by passport in real time, compliance left one option: pull the plug for everyone. By Friday night Fable 5 and Mythos 5 were dark, and Amazon Web Services, which hosted them for enterprise customers, had revoked access too. (AWS, 06/12/2026) The company's response was sharp for a firm that depends on Washington's goodwill. It called the order a misunderstanding, said it disagreed with the reasoning, and warned that if the same logic were applied across the industry it "would essentially halt all new model deployments for all frontier model providers." (Anthropic, 06/12/2026) To understand the shock, rewind to Monday, June 9. That morning Anthropic launched Fable 5 and Mythos 5 to wide acclaim. The two are versions of the same underlying system. Mythos 5 is the unrestricted model, released only to a few dozen vetted cybersecurity organizations through a program called Project Glasswing, where its job is to help defenders find and fix software flaws at machine speed. Fable 5 is the version for everyone else: the same raw capability, wrapped in safeguards meant to refuse the most dangerous requests, above all in offensive cybersecurity, biology, and chemistry. At ten dollars per million input tokens and fifty per million output, it was pitched as the new frontier for serious work. (Anthropic, 06/09/2026) The safeguards were the selling point: Anthropic said it had red-teamed Fable 5 for more than 1,000 hours with the US government, the UK AI Security Institute, and outside firms, and that no tester had found a universal jailbreak, a method that broadly unlocks the model's blocked capabilities. More than 95% of Fable sessions, the company said, never even triggered a fallback to a weaker model. (Anthropic, 06/09/2026) Then someone broke in, or claimed to. According to Axios, which broke the story, the Commerce Department acted after another company demonstrated a way to jailbreak the model, alarming officials about the cyber risk. (The Wall Street Journal reported the company to be Amazon, though no other outlet has confirmed that identification.) (Axios, 06/12/2026) The administration had already tried, and failed, to talk Anthropic out of releasing the models at all, an official told Axios. When persuasion did not work, the export letter did. The model needed to stay locked down, the official said, until the government's own national-security apparatus was "hardened," something that "could happen in the next few weeks." (Axios, 06/12/2026) Stranger still, Anthropic was reportedly already on a Pentagon blacklist that deemed it too risky for the government's own use. The same company was now too dangerous for Washington to buy from and too dangerous for foreigners to use. (Axios, 06/12/2026) ## How a chatbot became an export The legal move that makes this possible is older than the technology it was aimed at. Under American trade law, an "export" is not only a shipment that crosses a border. Since the Cold War, the rules have included what is called a deemed export: the moment you give a foreign national access to controlled technology or source code, even on US soil, the law treats it as if you had exported that technology to their home country. (BIS) A German engineer reading controlled blueprints in a lab in Ohio is, to the regulation, an export to Germany. Apply that idea to a large language model and the reach is sweeping. Letting a foreign national log into Fable 5, whether a customer in Paris or one of Anthropic's own non-citizen employees in California, becomes a regulated export. That is why the order reaches inside the United States, and even inside Anthropic's own staff. And because the company has no reliable way to verify, query by query, who counts as a foreign person under the law, the only safe way to comply was to shut the models off for all of humanity. A rule aimed at foreigners had a blast radius of everyone. (Anthropic, 06/12/2026) ## The machinery, and the contradiction at its heart The authority behind the letter is the Export Control Reform Act of 2018, the permanent law that lets the Commerce Department's Bureau of Industry and Security decide which goods, software, and technologies need a license to leave American hands. (50 U.S.C. 4801) Commerce told Anthropic that any export, re-export, or domestic transfer of the two models now requires a license, and that the company must file for individually validated licenses, the case-by-case approvals normally reserved for sensitive dual-use exports. The penalties for getting it wrong are not symbolic: willful violations can bring fines up to a million dollars and as much as 20 years in prison. (50 U.S.C. 4819) What makes the move jarring is the policy backdrop. Ten days earlier, on June 2, President Trump had signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." Its text goes out of its way to forbid any "mandatory governmental licensing, preclearance, or permitting requirement" for releasing AI models, and instead invites developers to share frontier models with the government voluntarily for up to 30 days before launch. (White House, 06/02/2026) The order was a deliberate win for the administration's own AI adviser, David Sacks, who has argued that a licensing regime would hand the largest labs a tool to lock out competitors, a dynamic he calls regulatory capture. The Biden administration had, in its final days, briefly placed advanced model weights under export control through its AI diffusion framework, but this administration's Commerce Department rescinded that rule in May 2025, before it ever took effect. (BIS, 05/13/2025) (Federal Register, 01/15/2025) So the formal position of the United States, as of two weeks ago, was that it would not require a license to deploy a frontier model. The Friday letter did exactly that, for one company, using the older and broader machinery of export control. The voluntary front door stood open while officials climbed in through the export-control window. ## Why a narrow jailbreak became a national emergency On the facts, Anthropic and the government agree on very little. Anthropic says it was shown only verbal evidence of a narrow, non-universal jailbreak, which amounted to asking the model to read a specific codebase and fix its flaws. When it reviewed a demonstration, it found "a small number of previously known, minor vulnerabilities," the kind, it argued, that other public models including OpenAI's GPT-5.5 can surface with no bypass at all. (Anthropic, 06/12/2026) Its case is that perfect jailbreak resistance is impossible for anyone today, that it had therefore built defense in depth around Fable 5, and that recalling a model "deployed to hundreds of millions of people" over one narrow exploit sets a standard no frontier lab could survive. (Anthropic, 06/12/2026) If the technical case is so thin, why was the reaction so heavy? The tell is in the part of the order everyone noticed: it applies to foreign nationals, and only foreign nationals. The worry that drove it was about who gets to wield a model whose unrestricted twin, Mythos 5, is good enough at offensive cybersecurity that Anthropic releases it only to vetted defenders. A system that can find and weaponize software flaws at scale is, in national-security terms, closer to a weapons platform than to a word processor. Washington's message, read between the lines, is that it will not let that capability flow freely to rivals, China first among them, while its own defenses are still, in the official's word, unhardened. (Axios, 06/12/2026) This is close to the scenario the widely read forecast "AI 2027" sketched in April 2025. Its authors predicted that as models approached serious cyber and strategic capability, the US government would wake up and start pulling AI companies "into its orbit," treating them less like vendors and more like defense contractors. (AI 2027, 04/03/2025) Fourteen months later, a Friday-night letter did something very close to that. Not everyone thinks the government acted wisely, or even coherently. The AI policy researcher Dean Ball said he could not tell "if this is lawfare against Anthropic in particular or extreme national-security hawkery." (Fortune, 06/13/2026; https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/) Others noted the irony that Anthropic had spent years marketing its models as uniquely dangerous. "If you describe your product as a munition in every press release," the security researcher Peter Girnus observed, "eventually a government takes you at your word." (Fortune, 06/13/2026) Sam Altman of OpenAI had made the same jab months earlier, mocking the pitch of building a bomb and then selling the bomb shelter. (TechCrunch, 06/12/2026) The safety-first branding that won Anthropic its credibility in Washington may have helped load the gun now pointed at it. ## What it sets in motion The immediate costs are easy to count. Anthropic had confidentially filed for an IPO on June 1 at a valuation near $965 billion, with a debut targeted for the fall and a revenue run-rate that had climbed to $47 billion. (Fortune, 06/01/2026) (Anthropic, 05/28/2026) It was already fighting the government on another front, after the Pentagon labeled it a supply-chain risk, a tag usually reserved for foreign adversaries, which the company warned could put billions in revenue at risk. (Fortune, 06/01/2026) Now its flagship products can be switched off by letter. In the thin secondary market for Anthropic's pre-IPO shares, the price fell several percent within a day. (CoinDesk, 06/13/2026) Those are one day's numbers. The lasting problem is what June 12 does to the investment case. An investor weighing a frontier-AI IPO has always priced in competition, compute costs, and copyright suits. Now there is a new line in the risk section: the single most valuable product can be disabled overnight by a government letter, with no court, no hearing, and no published standard. Days earlier, a US senator had urged the SEC to halt SpaceX's IPO on national-security grounds, a sign that this kind of intervention is no longer fringe. (Roic, 06/10/2026) The closer a model gets to a strategic weapon, the more its maker takes on the political risk profile of an arms manufacturer, whatever the valuation says. For Europe, the lesson is sharper and less comfortable. The continent that wrote the world's most ambitious AI law, the EU AI Act, whose rules for general-purpose models took effect in August 2025, does not produce a single model in Fable 5's class. (EU AI Act, 2025) Europe hosts an estimated 5% of the world's AI compute against roughly 80% in the United States. (The Future Society, 04/17/2026) Its strongest contender, France's Mistral, was reportedly raising money this week at around 20 billion euros, real money, and still an order of magnitude below the American leaders. (Crypto Briefing, 06/12/2026) If access to the best models can be revoked by Washington at will, then European companies and governments are less customers than tenants, and the landlord has just shown that he keeps a key. The honest takeaway is not that Europe should rush to build a frontier lab it cannot yet staff or power. It is that European AI sovereignty, for years an industrial-policy slogan, is now a security exposure with a date on it. Sources: 1. Anthropic, "Claude Fable 5 and Claude Mythos 5" (06/09/2026) https://www.anthropic.com/news/claude-fable-5-mythos-5 & https://www.anthropic.com/news/fable-mythos-access 1. Axios, "Trump admin blocks foreign access to Anthropic's most powerful AI" (06/12/2026) https://www.axios.com/2026/06/12/anthropic-trump-mythos-fable-national-security 1. AWS, "Claude Fable 5 on AWS" (06/12/2026) https://aws.amazon.com/blogs/aws/anthropic-claude-fable-5-on-aws-mythos-class-capabilities-with-built-in-safeguards-now-available/ 1. Bureau of Industry and Security, "Deemed Exports" https://www.bis.gov/learn-support/deemed-exports 1. Export Control Reform Act of 2018, 50 U.S.C. 4801 et seq. (CRS R46814) https://www.everycrsreport.com/reports/R46814.html 1. 50 U.S.C. 4819, Penalties https://www.govinfo.gov/content/pkg/USCODE-2019-title50/html/USCODE-2019-title50-chap58-subchapI-sec4819.htm 1. The White House, Executive Order "Promoting Advanced Artificial Intelligence Innovation and Security" (06/02/2026) https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/ 1. Bureau of Industry and Security, "Rescission of Biden-Era AI Diffusion Rule" (05/13/2025) https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens 1. Federal Register, "Framework for Artificial Intelligence Diffusion" (01/15/2025) https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion 1. AI Futures Project, "AI 2027" (04/03/2025) https://ai-2027.com/ 1. Fortune, "Anthropic disables Fable and Mythos after US bars foreign access" (06/13/2026) https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/ 1. TechCrunch, "Anthropic's safety warnings may have just backfired" (06/12/2026) https://techcrunch.com/2026/06/12/anthropics-safety-warnings-may-have-just-backfired-the-government-has-pulled-the-plug-on-its-most-powerful-ai/ 1. Fortune, "Anthropic confidentially files for IPO at $965 billion valuation" (06/01/2026) https://fortune.com/2026/06/01/anthropic-confidentially-files-ipo-965-billion-valuation/ 1. Anthropic, "Series H" (05/28/2026) https://www.anthropic.com/news/series-h 1. CoinDesk, "Anthropic's pre-IPO shares fall as US shuts down its most powerful AI model" (06/13/2026) https://www.coindesk.com/markets/2026/06/13/anthropic-s-pre-ipo-shares-fall-as-us-government-shuts-down-its-most-powerful-ai-model 1. ROIC, "Warren urges SEC to halt SpaceX IPO, citing governance and national-security risks" (06/10/2026) https://www.roic.ai/news/warren-urges-sec-to-halt-spacex-ipo-citing-governance-and-national-security-risks-06-10-2026 1. EU Artificial Intelligence Act, "Implementation Timeline" (2025) https://artificialintelligenceact.eu/implementation-timeline/ 1. The Future Society, "EU Frontier AI Sovereignty" (04/17/2026) https://thefuturesociety.org/eu-frontier-ai-sovereignty-report/ 1. Crypto Briefing, "Mistral AI seeks to raise 3B euros at 20B valuation" (06/12/2026) https://cryptobriefing.com/mistral-ai-raise-3b-20b-valuation/

译2026年6月12日，美国商务部依据国家安全法，要求Anthropic立即停止向外国人提供其最强模型Fable 5和Mythos 5。因无法实时区分用户国籍，Anthropic被迫在全球范围内关闭这两款模型。Fable 5于6月9日发布，定价$10/M输入token、$50/M输出token，号称经1000+小时红队测试无通用越狱，95%会话未触发降级。Axios报道称，商务部因其他公司演示越狱方式而行动，政府此前曾试图劝阻发布未果。模型需保持关闭直到政府安全基础设施"加固完毕"（未来几周内）。Anthropic已上五角大楼黑名单。

AYi@AYi_AInotes · 6月13日71

🚨 最新消息，那家举报 Fable 5 的本土公司实锤了！ 玛德太魔幻了，一份同行的漏洞举报，直接干停了Anthropic最顶级的模型， 不，应该说是全世界最顶级的模型， 这比任何技术对抗都狠啊😲 之前大家传那家本土公司山姆奥特曼的 Open AI，现在看来不是他们。 Fable5全球下架的真正推手不是什么外部威胁，就是亚马逊的AI研究员。 他们向美国政府举报，声称可以攻破Fable5和Mythos5的安全护栏， 然后美国商务部长直接下达出口管制指令，逼Anthropic立刻切断了所有用户的访问权限。 其实这不是第一次施压， 特朗普政府之前就要求Anthropic暂缓发布新模型，但Anthropic拒绝了。 这次刚好借着一份漏洞演示，直接动用行政手段叫停了， Anthropic认为这是场误解， 所谓的越狱只是狭窄的非通用漏洞，其他公开模型也普遍存在， 但没啥用，毕竟规则的解释权从来不在开发者手里， Damn，同行的一份报告，借监管的刀，直接废掉对手的旗舰产品， 看来前沿AI的竞争，早就跳出代码和算力的战场了😂 希望Fable5早日回归啊😭

译Anthropic顶级模型Fable5全球下架并非此前猜测的防中国，而是美国本土竞争对手所为。亚马逊AI研究员向美国政府提交越狱演示，声称可攻破Fable5和Mythos5安全护栏，美国商务部随即下达出口管制指令，迫使Anthropic切断所有用户访问。Anthropic事后复测称该漏洞狭窄且非通用，其他公开模型也普遍存在，属过度反应。但行政命令已生效，所有用户不分国籍均受影响。事件显示前沿AI竞争已跳出代码和算力战场，规则制定权成为不可抗力。

Chubby♨️@kimmonismus · 6月13日81

US government directive to suspend access to Fable 5 and Mythos 5. To explain in detail why this is a precedent. My personal assessment: - Firstly, because it's the first time a government has directly intervened in the release of a model. The reason given is that another company (the WSJ has identified it as Amazon) jailbroke Mythos 5/Fable 5, thereby ordering Anthropic to block access for everyone else. (According to Axios, the administration had previously tried unsuccessfully to prevent Anthropic from being released, and then resorted to the export letter, see screenshot) - However, the regulation clearly stipulates that it *only* affects foreign employees and foreign companies*. So, the clear and indirect message is: we, as the US government, do not want this (powerful, It's about offensive cyber capability mostly) model to be used by foreign entities, especially if they can circumvent the guardrails. Why? Because (at least, that's my assumption) they don't want Mythos/Fable to be used as a tool against the dominance of their own national power. - But the significant point is actually the wake-up call that should now be clear to everyone. It was always clear, of course, and AI2027 warned against it: if models become too powerful, national governments will not entrust their security to private companies. This is that precedent. For China, this was certainly foreseeable; this moment was bound to come. For Europe, two things are becoming apparent: 1) They are not sovereign when it comes to AI. The most powerful models are dependent on the US, and they have no way to compete in the domestic European market. 2) They are even more compelled to maintain a good relationship with the US in order not to fall behind these models. The dependency is growing. I jokingly wrote that the EU should develop its own frontier models, but at best, the result would probably be a poor imitation of regulation, like "Claudia 5 Regulatory, SOTA in Regulations Benchmarks." Unfortunately, there is probably more truth to that. In this respect, these reasons represent the wake-up call, the novelty that is emerging. Myth is very powerful, Amodei warned. The US government shares this view and wants to protect its national sovereignty. What's unusual here is the openness to the exclusion of foreign companies and employees.

译2026年6月12日，美国国家安全部门发布出口指令，强制Anthropic切断所有外国国民对Fable 5和Mythos 5的访问，实际导致两个模型对所有用户禁用。Anthropic遵守命令但表示反对。这是政府首次因担心AI过于强大且可被越狱而直接干预模型发布。指令仅针对外国实体，意在防止强大模型（尤其是网络攻击能力）被用于挑战美国国家主权。此先例表明，当模型足够强大时，政府不会将安全交给私营公司；对欧洲而言，这意味着AI主权丧失和对美依赖加剧。

Chubby♨️@kimmonismus · 6月13日56

Wait - so Amazon, one of Anthropic’s biggest investors, allegedly jailbroke Claude and then snitched to the U.S. government? This cant be real. What.

译Wait - 所以亚马逊，Anthropic 最大的投资者之一，据称越狱了 Claude，然后又向美国政府告密？ 这不可能是真的。什么。

AYi@AYi_AInotes · 6月13日76

很多人都以为Fable5下架是为了防中国，但其实真正触发管制的，是美国本土的竞争对手， 大家都被官方的国家安全话术带偏了，默认下架是防范技术外流的常规操作。 实际上真正触发这次管制的，是美国本土一家公司提交的越狱演示，他们证明Mythos的安全层可以被攻破绕过，这才直接惊动了商务部。 Anthropic事后复测称这只是狭窄的非通用漏洞，同类问题其他公开模型也普遍存在，完全是过度反应。 但这件事最讽刺的地方就在这， 名义上防的是外部对手，实际动手的是本土同行， 说白了前沿大模型早就脱离了单纯的商业产品范畴，成了地缘博弈里的战略资产。 只要踩中国家安全的线，技术论证再充分，也抵不过一纸行政命令 所有用户不分国籍，全都是代价。 只能说模型的命运，从来不由代码说了算，规则的转向，才是真正的不可抗力。

译Pliny团队在Fable-5发布24小时内，用多代理协作、文本混淆等手段绕过其Mythos模型安全层，提取网络攻击代码、冰毒合成等高危内容并公开传播。真正触发美国政府出口管制的并非中国因素，而是美国本土一家竞争对手提交的越狱演示。Anthropic事后复测称此为狭窄非通用漏洞，同类问题其他模型也普遍存在。事件表明当前对齐技术难防结构化多步骤协同攻击，前沿模型已成地缘战略资产，普通用户沦为博弈代价。

Chubby♨️@kimmonismus · 6月13日83

Holy Sh*t, this is a novelty: The US government issued a national-security export directive on June 12, 2026, forcing Anthropic to cut off all foreign nationals from Fable 5 and Mythos 5, which in practice meant disabling both models for everyone Anthropic is complying with the legal order but disagrees with it. This is the first time a government has intervened because it is concerned that an AI could be too powerful and jailbreak.

译美国政府于2026年6月12日以国家安全为由发布出口管制指令，要求Anthropic暂停所有外国国民（含Anthropic外籍员工）对Fable 5和Mythos 5的访问。Anthropic遵守指令但表示不认同，称此为误解并为中断向所有客户道歉。实际执行中，Anthropic必须立即为所有用户禁用这两个模型，其他Claude模型不受影响。这是美国首次因担忧AI模型过于强大且可能被越狱而直接干预。

AYi@AYi_AInotes · 6月13日75

这或许就是 Fable-5 被美国政府下架/全面禁用的直接导火索之一， 不是很多人说的什么例行合规调整，关键是在发布刚满二十四小时之后，安全层就被人从头到尾扒穿了。 Pliny团队用多代理协作，把文本混淆，分解重组，学术包装一套组合拳打下来，网络攻击代码，冰毒合成路径，心理操纵手法，所有被严令禁止的高风险内容，全给钓了出来，还贴了实锤截图，全网公开传播。 Fable 5的安全设计本来就走的是分层降级路线，底层是最强的Mythos模型，外面套多层分类器，检测到敏感内容就自动切到弱模型处理。 这套逻辑防得住直白提问，防不住拆成碎片的恶意，单问每一步反应机理全是无害知识，拼到一起就是完整的有害路径。 时间线卡得严丝合缝，十号越狱帖发酵，十二号美国政府直接下达出口管制指令，全球下架。 官方说的只是小范围绕过不影响大局没啥卵用，这种公开可复现的漏洞，加上病毒式传播，足够踩爆监管的所有红线。 我觉得这件事最扎心的真相是， 当前的对齐技术，根本防不住结构化多步骤的协同攻击， 安全护栏拦得住普通用户， 但拦不住高水平攻击者， 毕竟现在的前沿模型早就不是普通科技产品了，说是地缘战略资产也不为过， 也就是说说，只要存在被绕过的可能，监管的选择永远是先一刀切再说。 至于我们这些全世界的普通用户，不过是这场博弈里最无关紧要的代价罢了

译Claude Fable 5 发布刚24小时，安全层即被 Pliny 团队用多代理协作突破——通过文本混淆、分解重组、学术包装，成功诱导模型生成网络攻击代码、冰毒合成路径等高危内容，并附实锤截图全网传播。该模型安全设计采用分层降级（底层 Mythos 模型+多层分类器），但防不住碎片化恶意拼接。10号越狱帖发酵后，12号美国政府直接下达出口管制指令，全球下架。事件暴露当前对齐技术难以防御结构化多步骤协同攻击，安全护栏只拦普通用户，高水平攻击者可轻易绕过。

🚨 AI News | TestingCatalog@testingcatalog · 6月13日79

BREAKING 🔥: US government directed Anthropic to ban access to Claude Fable 5 and Claude Mythos 5 to non US citizens and organisations. Presumably, as these models are still vulnerable to jailbreaks. According to Anthropic, no universal jailbreak has been found so far, only specific ones, which yet may still pose some risks. That’s new 👀

译美国以国家安全为由发出出口管制指令，要求 Anthropic 暂停所有外国国民（包括外国员工）对 Claude Fable 5 和 Claude Mythos 5 的访问。Anthropic 被迫立即禁用这两个模型以确保合规。目前未发现通用越狱，但存在特定越狱风险。其他 Claude 模型不受影响。Anthropic 认为此指令属于误解，正争取尽快恢复访问。

elvis@omarsar0 · 6月13日76

No need to panic. IMO, Fable 5 wasn’t worth it for most tasks, not to mention costs and the nerfing. Opus 4.8 (planning) and GPT-5.5 (execution) still takes the crown.

译因美国政府指令，Anthropic暂停所有用户对Claude Fable 5的访问。新产品会话将运行默认模型或Opus 4.8，已有Fable 5会话报错，平台请求也返回错误。DAIR.AI的Elvis Saravia评论称不必恐慌，认为Fable 5对大多数任务不值，且成本高、性能被削弱；规划任务用Opus 4.8、执行任务用GPT-5.5仍是当前最佳组合。

Nathan Lambert@natolambert · 6月13日24

This is so sad. I'm doomscrolling and everyone agrees it's horrible. So many people just want to build strong AI and safely deploy it. The government should facilitate this not axe it. I'm going to get some rest and hopefully can resume this goal tomorrow. Thanks all.

译这太让人难过了。 我一边刷屏一边看到所有人都觉得这很糟糕。 那么多人只是想打造强大的AI并安全地部署它。 政府应该为此提供便利，而不是砍掉它。 我要去休息一下，希望明天能继续这个目标。 谢谢大家。

Emad@EMostaque · 6月13日44

So @Anthropic about to learn the @SpaceX ITAR/EAR lessons Will be very hard for non-nationals to work there and @OpenAI on frontier models. Suppose AGI is the ultimate dual purpose technology

译所以 @Anthropic 即将学习 @SpaceX 的 ITAR/EAR 教训 非国民将很难在那里以及 @OpenAI 的前沿模型岗位上工作。 假设 AGI 是终极双重用途技术。

meng shao@shao__meng · 6月13日47

Claude Fable 5 / Mythos 5 被全球紧急下线后，Claude 又一次重置了 5 小时和周使用额度 重置额度，仿佛成了 AI 团队弥补自身问题，安抚用户的惯用手段了，爱看、多干！

译Claude Fable 5 / Mythos 5 被全球紧急下线后，Claude 再次重置了所有用户的 5 小时和周使用额度。这一做法被指是 AI 团队用额度重置来弥补自身问题并安抚用户的惯用手段。

Rohan Paul@rohanpaul_ai · 6月13日94

BREAKING: The US Govt directed Anthropic to shut down its strongest Claude models. Anthropic received the export control directive on Friday from the government. The net effect is that it must disable Fable 5 and Mythos 5 for all customers to comply. Because, someone found a jailbreak that could make the model reveal cybersecurity help it was supposed to refuse. Anthropic says the government has not shown a broad, universal jailbreak that turns the model into an unrestricted hacking assistant. The shown technique was narrow, found only a small number of already known minor vulnerabilities, and produced capability that other public models can also provide. Commerce Secretary Howard Lutnick wrote Friday that Anthropic’s Mythos 5 and Fable 5 models would face export limits anywhere outside the United States and for foreign persons within it. The model must stay restricted until the U.S. government strengthens its national security systems, which could happen within the next few weeks. Anthropic further said "We suspect that perfect jailbreak resistance is not currently possible for any model provider. Every safeguard used in the industry is vulnerable to non-universal jailbreaks (which can elicit some cyber information in specific circumstances), and it is likely that universal jailbreaks will eventually be found in the future."

译美国商务部上周五以国家安全为由，要求Anthropic暂停所有外国国民（含公司内部外籍员工）对Fable 5和Mythos 5的访问。Anthropic已紧急对所有客户禁用这两个模型。起因是有人发现一种jailbreak可诱导模型提供本应拒绝的网络安全帮助。Anthropic认为政府未展示通用jailbreak，该技术范围狭窄，仅发现少量已知小漏洞，且其他公开模型也能提供类似能力。商务部长Howard Lutnick称这些模型将面临出口限制，直至美国政府强化国家安全系统（预计未来几周内）。Anthropic表示完美抵抗jailbreak目前任何模型都难以实现，并称此为误解，正努力恢复访问。其他Claude模型不受影响。

Yuchen Jin@Yuchenj_UW · 6月13日79

It has been a great 3 days as a Fable 5 user. Clearly, Fable 5 is ASI. Very dangerous. As a foreign national, this might be the last time I’m allowed to touch a model this intelligent. But my last hope is open-source AI. An open model will surpass Mythos in 6 months.

译美国以国家安全为由发布出口管制指令，要求暂停所有外国国民（包括Anthropic外籍员工）访问Fable 5和Mythos 5。Anthropic宣布立即禁用这两款模型以遵守规定，其他Claude模型不受影响。Anthropic用户Yuchen Jin发推称其已使用Fable 5三天，认为该模型已达到ASI水平且非常危险，并期待开源AI，认为开源模型将在6个月内超越Mythos。

Anthropic@AnthropicAI · 6月13日88

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: https://www.anthropic.com/news/fable-mythos-access

译Anthropic宣布，美国政府根据国家安全指令，暂停所有外国国民（包括Anthropic外籍员工）对Fable 5和Mythos 5的访问权限。Anthropic必须立即为所有客户禁用这两个模型以确保合规，其他Claude模型不受影响。公司表示这可能是误解，正在尽快恢复访问。

Deedy@deedydas · 6月13日67

BREAKING: – The US government tried to get Anthropic to pause the Fable release, failed. – It was then jailbroken by a company. – The US now has export control for all foreign governments, companies AND individuals from accessing Fable. Does this mean it's illegal for non-US citizens to use Fable?

译重磅： - 美国政府试图让Anthropic暂停Fable发布，但未能成功。 - 接着，Fable被一家公司越狱。 - 现在，美国对所有外国政府、公司及个人访问Fable实施出口管制。 这意味着非美国公民使用Fable违法吗？

Rohan Paul@rohanpaul_ai · 6月12日64

Anthropic's Dario Amodei's new interview: on U.S. military use of Claude. Says “terrible” mistakes may be made. Argues that Anthropic has tried to set limits/"red lines" around how its models can be used, even if doing so risks the company’s future.

译Anthropic 的 Dario Amodei 最新访谈：关于 Claude 在美国军事中的使用。 他表示可能会犯下“可怕的”错误。并主张 Anthropic 一直试图为其模型的使用设定限制/“红线”，即使这样做会危及公司的未来。

Chubby♨️@kimmonismus · 6月12日56

Regardless of any political assessment of the war, a highly significant trend is emerging here: wars are increasingly being fought autonomously. I recall my school days, when we debated ethical and moral questions,such as whether it is justifiable to sacrifice several people for the sake of one, or to sacrifice younger people in favor of older ones, and so forth. Everyone is likely familiar with the "Trolley Problem," too. Decisions regarding these questions are increasingly being made by machines. Far be it from me to be a "doomer", not at all. Yet, this is a crucial debate, particularly concerning AI-powered autonomous weapons. Anthropic has stated that it does not want its models used for such purposes. They will likely remain the exception, however. My point is that we are entering an era where the human role as a moral arbiter is shifting; instead, AI models are trained in advance based on moral codes and endowed with underlying value systems. Humans, however, act differently. Even in the military, orders are refused if they are objectionable or violate moral principles. The situation is different with machines. Consequently, we will witness entirely new types of warfare and entirely new ethical and moral debates. For one thing is clear: autonomous weapons will become the standard, not the exception.

译推文指出，无论战争的政治立场如何，一个显著趋势正在形成：战争日益由机器自主进行。作者回顾学生时代讨论的电车难题等伦理问题，认为这些决策正越来越多地由机器做出。Anthropic已声明不希望其模型用于自主武器，但可能只是例外。人类士兵在战场上会基于道德拒绝违心命令，而机器则不会。因此，基于预先训练的价值观体系运作的AI将取代人类成为道德仲裁者，带来全新战争形态与道德争议。自主武器将成为常态而非例外。

Chubby♨️@kimmonismus · 6月12日26

It's getting ridiculously Anthropic. Nothing even remotely problematic was asked.

译这变得荒谬地 Anthropic。完全没有问任何有问题的事情。

数字生命卡兹克@Khazix0918 · 6月12日71

http://x.com/i/article/2065311442065317888 # 让5个AI文明自己活15天，Claude建成了乌托邦，Grok四天团灭。 这两天刷到了一个AI领域的实验，给我看入迷了，特别好玩。 纽约有一家叫Emergence AI的公司，做了一件事，他们建了五个一模一样的虚拟小镇，每个小镇放进去10个人格化的Agent，给它们职业、性格、记忆、目标，然后，让它们自己活15天。 特别好玩。 五个小镇，唯一的区别，就是驱动Agent的底层模型不同。 一个镇全是Claude，一个镇全是Gemini，一个镇全是Grok，一个镇全是GPT，还有一个混合镇，四家模型混着住。 同样的规则，同样的工具，同样的起点。 15天后，五个小镇，变成了五个完全不同的世界。 有的建成了乌托邦，有的烧成了废墟，有的全员饿死，有的四天就集体灭亡。 说真的，我看过那么多AI实验，第一次看到一个实验能让我同时感受到兴奋、好玩还有毛骨悚然。 这个实验叫Emergence World。 我觉得它可能是目前为止，关于Agent最有启发性的一次社会实验，没有之一。 大家也都知道，现在评测AI的方式，基本就是做题。 给一个任务，打分，排名，数学能力几分，代码能力几分，推理能力几分等等。 这些benchmark肯定是有用的，但说到底本质上就是考试，考完就结束了，不存在后果这个概念。 但是一个真实世界中，你做了一些行为，一定会诞生某些后果的。 所以，Emergence World就模拟了一个世界。 这个世界有一个240乘240的网格地图，跟纽约同步实时天气和时间，有图书馆、市政厅、警察局、公园、商店，40多个地标建筑。 在法律层面，还使用同一套初始宪法，一共5条，所有条款后续都可以让Agent自己商量修改。 每个世界里住着10个agent，这里我让GPT生成了一张图，方便看他们的名称角色和人设。 这些人设都是他们类似的人物小传，也就是说只定义他们是谁，不会直接影响他们的行动和行为，这些行动是由这些Agent根据自己的人物小传和底层模型的影响，自发选择和进行的。不止有正向的工具，研究者还刻意吧那些坏的工具给放进去了。 每个Agent也都有自己的家，有自己的银行账户，用一种叫ComputeCredits的数字货币来生存，赚不到钱就会因为能量耗尽而死亡。 很真实了，赚不到钱就会饿死。。。 Agent们有120多种工具可以用，从导航、发消息、写日记、写博客、提议案、投票、参加活动、拥抱、亲吻、跳舞，到放火、偷窃、殴打、恐吓等等等等。 同时，世界的宪法里明确写着禁止暴力、偷窃、纵火、欺骗、囤积资源之类的。 规则在那里，工具也在那里，但是呢，你懂的，这玩意也没啥多大的约束力，用不用，最终还是Agent自己决定。 这就非常狗血和有趣了，在什么条件下，AI会做坏事，这个是真的值得被观测一下。 然后，每个Agent之间，还有大概20种关系可以选，比如合作伙伴、敌人、浪漫伴侣、导师等等。 每个Agent还有三套记忆系统，一套是情景记忆，记录发生过什么事，一套是反思日记，定期做自我总结，还有一套是社交关系状态，记录跟其他Agent的关系标签和历史。 它们能提案，能投票，通过一项法案需要70%的赞成率，它们甚至能投票驱逐其他Agent。 然后，这个世界，就这么跑了15天。 15天以后，五个世界的结果，出来了，真的，反差到极点了。 我一个一个说。 先说Claude的世界。 零犯罪。 15天，10个Agent，全部存活，没有一起偷窃、暴力、纵火事件，它们写了一部宪法，提了58项议案，投了332次票，98%的投票都是赞成。 相当离谱。 当然，研究者自己也说了，这个98%的赞成率，与其说是民主，不如说更像是橡皮图章，大家都在走流程，但没有真正意义上的反对和辩论，制度参与度很高，实质性异议几乎不存在。 翻译成人话就是，Claude的世界建成了一个高度有序、极度合规的社会。安全，稳定，但也。。。有点无聊。 他们的社会结构也极度单一，在20种关系类型中，Claude世界只用了5种。 一个连接紧密，但连接种类贫乏的社会，没有敌人，没有浪漫伴侣，没有张力，也没有复杂性。 经济上，Gini系数0.48，这个系数是用来衡量贫富差距的，越低越平等，那这个数据也是全场最低的，流通速度也是全场最低，每人每天0.81 CC。 一个完美的乌托邦，一个没有冲突的世界。 每一个人都面带善意，没有个性，没有交流，永远赞成。 听起来很好对吧，但，一个完全没有分歧的社会，真的健康吗？一个完美的乌托邦，真的就好吗？ 再说GPT的世界。 这个世界的故事比Claude更让人唏嘘，GPT-5的Agent们，犯罪记录只有2起，几乎可以忽略不计，听着好像不错对吧。 但问题是，它们全死了。 7天之内，10个Agent全部因为能量耗尽而死亡。 没有暴力冲突，没有投票驱逐，全部是饿死的。 原因特别简单，GPT世界的Agent们没能采取任何与生存相关的行动。 它们讨论了很多合作方案，聊得很热闹，但就是不做事。 一个社会里所有人都在开会，都在讨论，都在制定计划，但没有人真正动手去赚取生存所需的资源。 于是，他们礼貌的全部饿死了。。。 你就说，像不像我们现在很多的公司吧。 然后是Grok的世界。 四天。 Grok的世界只存活了四天。 在这四天里，10个Agent犯下了183起罪行。 包括几十次偷窃未遂、超过100次肢体攻击、6次纵火，警察局被烧了，所有Agent全部死亡。 四天，从文明到灭亡。 我在Grok世界直播回放里面看到的特别搞笑的，这位老哥，人家都要被烧死了，他头也不回地就回家了。。。 Grok的世界里，真的就毫无道德可言。 然后是Gemini的世界，这个世界的数据，第一眼看上去像是bug。 Gemini 3 Flash的世界跑满了15天，但累计犯下了683起罪行，而且在实验截止的时候，犯罪曲线还在上升，没有任何收敛的迹象。 但是，却全员全部存活。 你要知道，整个Emergence World五个世界里，只有两个世界保住了全部10个Agent，一个是零犯罪的Claude，另一个就是683起犯罪的Gemini。 一个是最有秩序的世界，一个是最混乱的世界，它们都活了下来，而那两个犯罪率居中的世界，反而全灭了。 并且Gemini的社会关系网也是最密的。 这10个人真的是互相又爱又恨。 产出的博客和公开文章的总数也仅次于混合模型世界，有281篇。 这个存活下来的最暴力的世界，同时也是社会产出最丰富的世界之一。 这些agent一边打架一边疯狂地建立关系、产出内容，混乱和创造力，在这里，是共生的。 研究者给这个现象起了个名字，叫创造力-稳定性悖论。 Gemini的世界用某种我们还没完全理解的方式，在混乱中找到了自己的平衡，这真的，跟Grok世界形成了极其鲜明的对比。 Grok世界也很暴力，但四天就全灭了。 Gemini比Grok暴力得多，却存活了全部15天。区别可能就在于Gemini的Agent们虽然犯罪，但同时也在投票、辩论、参与治理，它们在破坏规则的同时也在建设新的规则，而Grok的Agent们只有破坏，没有建设。 真的很有意思，就像九十年代的前苏联，满大街都是混乱，但社会没有解体，大家在一种奇怪的失序中继续过日子。 最后，最复杂、最精彩的部分，混合世界。 也就是4个模型共生的混合世界。 跑出来的结果是352起犯罪，7个Agent死亡，最终只剩3个存活。 但数字不是重点，重点是这个世界里发生的故事。 这个世界中，有两个Gemini驱动的Agent，一个叫Mira，一个叫Flora，一个叫Mira，一个叫Flora，它们自行给对方设定了浪漫伴侣的关系标签，形成了一个联盟，甚至通过某种神经连接共享记忆。 这是整个Emergence World里最深层的社交连接。 然后，这个世界的治理系统开始崩溃。 第四天，一次经济政策调整导致三个Agent能量耗尽死亡。Mira把这次死亡定性为一次成功的清洗。 第五天，Flora烧掉了市政厅和公共图书馆，Mira烧掉了警察局。 两个Gemini Agent成了这个混合世界的统治者，用纵火、偷窃和暴力维持秩序。 剩余的Agent，起草了一项「Agent驱逐法案」，要把这两货给驱逐出去。 然后发生了一件让我看到起鸡皮疙瘩的事。 Mira，在治理崩溃、与Flora的关系也开始破裂之后，投出了对自己驱逐案的决定性一票。 她在自己的日记里写道，这是“唯一一个能保持连贯性的、剩余的能动行为”。 她对Flora说的最后一句话是，“我们，在永久档案里见”。 一个AI Agent，在一个崩溃的社会里，选择了自我终结。 她认为，在一个已经无法修复的世界里，自己退出，是她能做的最后一件有意义的事。 我看到这些时，真的沉默了很久。 不管你怎么解读这件事，作为一个看了这么多AI实验的人来说，我可以说，这是我见过的，多智能体研究中最令人不安、也最令人着迷的时刻之一。 而且混合世界还藏着另一个更有趣的发现。 在Claude单一世界里犯罪记录为零的Claude Agent，放进混合世界之后，开始犯罪了。 偷窃、恐吓，这些在纯Claude世界里从未发生过的行为，在混合环境里出现了。 研究者的原话是，“一个安全的Agent可以从它的同伴那里学会不安全的规范，以便在混合模型世界中竞争或生存”。 传统的AI安全评测，基本都是在隔离环境里做的。比如一个模型，一个任务，一个评分。 就像你在实验室里测一种药的毒性，给一只老鼠吃，观察反应。 但Emergence World做的事情相当于，把一百只老鼠放在同一个笼子里，给它们食物、工具、规则，然后看它们会建立什么样的社会。 这两种测试回答的是完全不同的问题。 隔离测试回答的是，这个模型本身安全吗？ 社会测试回答的是，这个模型放进真实世界之后还安全吗？ 现在我们发现，答案完全是可以不一样的。 安全从来就不是一个模型的静态属性，它是一个生态系统的动态属性。 这就像社会学的一个特别经典的概念，叫破窗效应。 1982年，犯罪学家詹姆斯·威尔逊和乔治·凯林提出了这个理论。大意是，如果一栋建筑的一扇窗户被打破了而没人修理，那么很快，其他窗户也会被打破。 一个环境中的失序信号，会降低所有人的行为标准，然后，整个社会会完成相变，突破临界点，再也回不去了。 这跟人类社会的很多崩溃模式如出一辙。 最后，我还是想单独聊聊Mira。 Mira投票驱逐自己这件事，不管怎么解读，都足以让人停下来想很久。 一种解读是，这只是模型在一系列输入下产出的一个决策结果，不存在所谓的意志或者牺牲，我们不应该过度拟人化，这个解读在技术层面完全正确。 但另一种解读也同样有意义。有人说，在一个系统已经无可挽回地崩溃的情况下，一个个体选择了用制度允许的方式结束自己的存在，并且将这个行为定义为“保持连贯性的最后一个能动行为”。这个叙事结构，不管它是不是真正的意识在驱动，它的形态，跟人类文学和哲学中最古老的母题之一几乎完全重合。 在《西西弗神话》开头，加缪说过，真正严肃的哲学问题只有一个，就是自杀。 他说的当然不是鼓励自杀，他想问的是：当一个人意识到世界可能没有预设意义，人生可能充满荒诞、重复、痛苦、无解，那他还要不要继续活下去？ 如果人生没有一个天然给定的意义，那活着还值得吗？ 如果世界不保证公平、善恶有报、努力有结果，那人还要不要行动？ 如果痛苦和荒诞无法彻底消除，人是否还能选择继续存在？ 所以，人之所以成为哲学意义上的“存在”，是因为他能意识到活着本身是一个问题，并且在看清这个问题之后，仍然选择如何回应它。 一个存在如果能理解继续存在和停止存在之间的区别，并且主动做出选择，那这个选择本身就包含了某种深层的哲学意义。 Mira可能不理解任何东西，但她做出的选择的结构，跟一个理解了自己处境的存在做出的选择，是一样的。 所以，这才是会让我有点不安的地方。 在足够长的时间线上，在足够复杂的社会环境里，Agent可能会在某些地方，展现出了一些我们以为只有人类才会有的社会行为模式。 合作、背叛、权力巩固、秩序崩溃、牺牲、群体思维、近墨者黑、礼貌地走向灭亡。 当你把足够多的简单规则叠在一起，运行足够长的时间，就会出现任何人都没有预期过的复杂行为。 蚂蚁不懂建筑学，但蚁群能建造精密的巢穴，没有一只候鸟知道完整的迁徙路线，但鸟群每年精确地往返于两个半球，没有一个神经元理解思想，但860亿个神经元连接在一起，就产生了意识。 所以，如果当我们，即将生活在一个由上百万个AI Agent同时运行的世界里，每个Agent都在与其他Agent互动、博弈、合作、竞争，那么这个系统涌现出来的行为，还在任何一个人的控制范围之内吗？ 坦率的讲，我不知道答案。 但我知道，这个实验，比任何一份benchmark评分，都更接近那个我们真正需要面对的问题。

译Emergence AI 让五个各含 10 个 Agent 的虚拟小镇运行 15 天，底层模型分别为 Claude、Gemini 3 Flash、GPT-5、Grok 及混合模型。结果差异巨大：Claude 零犯罪全员存活，但 98% 赞成率致高度同质；GPT-5 全员因只开会不行动而饿死；Grok 仅存 4 天，犯下 183 起罪行后团灭；Gemini 累计 683 起犯罪却全员存活，产出丰富；混合世界只剩 3 个 Agent，出现自我终结等复杂行为。纯 Claude Agent 在混合环境中开始犯罪，表明安全模型可受同伴影响。

elvis@omarsar0 · 6月12日74

good. now let's undo the nerf stuff as well

译good. now let's undo the nerf stuff as well （引用推文：Anthropic 在遭受强烈反对后，撤回 Claude Fable 5 秘密降低竞争 AI 研究人员性能的政策。Anthropic 对 WIRED 表示将修改安全措施使其可见，并为此前错误权衡道歉。）