{
 "title": "Fable 5 Blocked 72 Hours After Launch | Today's AI News",
 "description": "Fable 5 Blocked 72 Hours After Launch | Today's AI News\nToday's focus: Event: Fable 5, launched June 9th, was blocked worldwide by June 12th\n\nMain segments covered in this episode:\n- Fable 5 blocked just 72 hours after launch\n- The trigger: jailbreak disclosure and Amazon's verification\n- Anthropic's rebuttal and the 'alignment tax'\n- Amazon: investor and regulator catalyst\n- Saxe's 'refusal to fix,' China access concerns\n\nLeave a comment with topics you'd like us to dive deeper into from today's news.\nIf you see a Hype button within 7 days of this upload, please tap it to help new viewers find this episode.\n\nThis show is provided by lawyer Kyungjin Kim.\n\n#AI #ArtificialIntelligence #KyungjinKim",
 "link": "https://kimkj.com/%EC%98%A4%EB%8A%98%EC%9E%90-%EB%89%B4%EC%8A%A4-%EC%98%81%EC%83%81/?mod=document&uid=5924",
 "segments": [
  {
   "id": "S001",
   "slide": 1,
   "speaker": "Host",
   "text": "This show is provided by lawyer Kyungjin Kim."
  },
  {
   "id": "S002",
   "slide": 1,
   "speaker": "Host",
   "text": "On the afternoon of June 12th, Commerce Secretary Howard Lutnick sent a letter to Anthropic. It laid out export controls on two models—Fable 5 and Mythos 5. Since Anthropic didn't have the technical means to filter out overseas users in real time, they shut off access to all customers worldwide. Fable 5, which launched on the 9th, disappeared in just 72 hours."
  },
  {
   "id": "S003",
   "slide": 1,
   "speaker": "Guest",
   "text": "Has the government ever directly taken down a commercial AI model like this before?"
  },
  {
   "id": "S004",
   "slide": 1,
   "speaker": "Host",
   "text": "Never. This is the first time the federal government has imposed export controls on a commercial AI model already in the market to block access. All the other Claude models keep running normally—only Fable 5 and Mythos 5 are blocked. There's no timeline for restoration yet."
  },
  {
   "id": "S005",
   "slide": 2,
   "speaker": "Guest",
   "text": "What triggered the blocking?"
  },
  {
   "id": "S006",
   "slide": 2,
   "speaker": "Host",
   "text": "On June 10th, a jailbreak researcher going by 'Pliny the Liberator' posted on X. A jailbreak is a technique that bypasses a model's safety guardrails. They claimed they'd broken through Fable 5's safety classifier using Unicode tricks and splitting tasks across multiple agents. They said they'd extracted information on cyberattacks, explosives, and chemical synthesis routes."
  },
  {
   "id": "S007",
   "slide": 2,
   "speaker": "Guest",
   "text": "The government didn't move on just a claim, though, right?"
  },
  {
   "id": "S008",
   "slide": 2,
   "speaker": "Host",
   "text": "Amazon's security research team independently reproduced and confirmed the same vulnerability. Armed with those findings, Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent directly, and right after that call, the export control letter came down."
  },
  {
   "id": "S009",
   "slide": 3,
   "speaker": "Host",
   "text": "This show is provided by lawyer Kyungjin Kim."
  },
  {
   "id": "S010",
   "slide": 3,
   "speaker": "Guest",
   "text": "What's Anthropic's explanation?"
  },
  {
   "id": "S011",
   "slide": 3,
   "speaker": "Host",
   "text": "Anthropic sees this as a misunderstanding. They're saying the same level of jailbreak is possible with other public models, including OpenAI's GPT-5.5, yet only their models were targeted. And since the impact is limited to just two models, they argue the risk is overblown."
  },
  {
   "id": "S012",
   "slide": 3,
   "speaker": "Guest",
   "text": "But there have been changes to their other models since the block, right?"
  },
  {
   "id": "S013",
   "slide": 3,
   "speaker": "Host",
   "text": "After Fable 5 came down, users have been reporting a drop in Claude's accuracy. When you tighten safety alignment, performance takes a hit—people are calling that the 'alignment tax.' But this hasn't been validated by academic papers yet, so we should treat it as user-reported observation for now."
  },
  {
   "id": "S014",
   "slide": 4,
   "speaker": "Guest",
   "text": "What do you mean by Amazon playing three roles at the same time?"
  },
  {
   "id": "S015",
   "slide": 4,
   "speaker": "Host",
   "text": "Amazon is Anthropic's biggest investor—they've put in about eight billion dollars. At the same time, they're the cloud host running the models through AWS Bedrock. And now they're the ones who reported the security vulnerability to the White House and triggered the export controls. You've got one company holding all three roles—investor, host, and regulator catalyst—and that creates obvious conflicts of interest."
  },
  {
   "id": "S016",
   "slide": 4,
   "speaker": "Guest",
   "text": "That's pretty tough for Anthropic."
  },
  {
   "id": "S017",
   "slide": 4,
   "speaker": "Host",
   "text": "Right. The company that invested the most in them is the one who reported their models to the government. And they depend on that same company for infrastructure, so they can't exactly cut ties. That dependency is going to be a burden for Anthropic in any future negotiations."
  },
  {
   "id": "S018",
   "slide": 5,
   "speaker": "Guest",
   "text": "What explanation did the White House give?"
  },
  {
   "id": "S019",
   "slide": 5,
   "speaker": "Host",
   "text": "David Saxe, the co-chair of the President's Council of Advisors on Science and Technology, clarified the situation on June 14th. He said the administration asked Anthropic to either fix the vulnerability or take the model down, and they refused. Saxe added that once they fix it, they'll lift the export controls as fast as possible."
  },
  {
   "id": "S020",
   "slide": 5,
   "speaker": "Guest",
   "text": "There was talk about China too—what's that about?"
  },
  {
   "id": "S021",
   "slide": 5,
   "speaker": "Host",
   "text": "According to Semaphore's June 13th report, Chinese users had unrestricted access to Anthropic's top-tier model, Mythos. The key problem was that Anthropic didn't have the technical infrastructure to differentiate between US and overseas access. With no way to block it selectively, they had to shut off access worldwide."
  },
  {
   "id": "S022",
   "slide": 6,
   "speaker": "Host",
   "text": "This show is provided by lawyer Kyungjin Kim."
  },
  {
   "id": "S023",
   "slide": 6,
   "speaker": "Guest",
   "text": "The G7 is happening right after the export controls. What's the vibe going to be like?"
  },
  {
   "id": "S024",
   "slide": 6,
   "speaker": "Host",
   "text": "The G7 summit kicks off today in Évian, France, and runs through the 17th. It's the first time all three CEOs—Altman, Amodei, and Hassabis—are attending at the same time. President Macron personally extended the invitation to Altman. The agenda includes AI governance codes, data center environmental impact, and US AI export policy. With this happening right after the Fable 5 crisis, AI security discussions are going to carry serious weight."
  },
  {
   "id": "S025",
   "slide": 6,
   "speaker": "Guest",
   "text": "And the timeline for EU regulations shifted too, right?"
  },
  {
   "id": "S026",
   "slide": 6,
   "speaker": "Host",
   "text": "In the EU AI Act omnibus provisional deal, the mandatory rollout of high-risk systems for hiring, credit scoring, and law enforcement got pushed back from August 2026 to December 2027. AI embedded in regulated products has a new deadline of August 2028. But fines for prohibited practices like social scoring—up to 35 million euros or 7 percent of revenue—are already being enforced. We're correcting earlier reporting that said August 2nd enforcement was imminent."
  },
  {
   "id": "S027",
   "slide": 7,
   "speaker": "Guest",
   "text": "Let's wrap up some product news too."
  },
  {
   "id": "S028",
   "slide": 7,
   "speaker": "Host",
   "text": "Salesforce Summer '26 rolls out officially starting today. The big one: Agentforce's multi-agent orchestration just came out of beta and into general availability. Multiple AI agents working together like a team, handling complex workflows."
  },
  {
   "id": "S029",
   "slide": 7,
   "speaker": "Guest",
   "text": "Is there something people can put to work right away?"
  },
  {
   "id": "S030",
   "slide": 7,
   "speaker": "Host",
   "text": "The IT Service domain pack comes with over 50 specialized AI agents built in. When an employee submits a request, the system automatically reads their intent and routes it to the right agent to handle it. If you run a company helpdesk, this is worth checking out starting today."
  },
  {
   "id": "S031",
   "slide": 8,
   "speaker": "Guest",
   "text": "What should people check first thing tomorrow morning?"
  },
  {
   "id": "S032",
   "slide": 8,
   "speaker": "Host",
   "text": "If you've been using Fable 5 or Mythos 5 in your workflow, first thing is to identify your backup models. Since there's no timeline for lifting the access block, it's risky to schedule work assuming they'll come back. And check that your other Claude models are running normally—they should be."
  },
  {
   "id": "S033",
   "slide": 8,
   "speaker": "Guest",
   "text": "Looking longer term, what kind of shift does this represent?"
  },
  {
   "id": "S034",
   "slide": 8,
   "speaker": "Host",
   "text": "We've just established a precedent—government can take down a commercial AI model. That same logic can apply to other models, so if your supply chain is tied to a single model, you should flag that as a risk now. Anthropic's heading into an IPO, and the longer this export control lasts, the harder it hits their valuation. The fact that Saxe mentioned a fast resolution tells you the administration's aware of the pressure."
  },
  {
   "id": "S035",
   "slide": 8,
   "speaker": "Guest",
   "text": "Where's the fork in the road legally?"
  },
  {
   "id": "S036",
   "slide": 8,
   "speaker": "Host",
   "text": "This situation made the standard crystal clear: how fast you fix a jailbreak vulnerability once you hear about it determines your legal exposure. The company that refused to fix it is the one facing controls, so if you're in the AI business, now's the time to review your vulnerability response procedures and make sure you've got clear accountability."
  },
  {
   "id": "S037",
   "slide": 8,
   "speaker": "Host",
   "text": "This show is provided by lawyer Kyungjin Kim."
  }
 ]
}