US government warned Anthropic that Fable 5 had been jailbroken, but firm 'refused' to fix before US implemented export controls — Anthropic defended its decision by saying the jailbreak 'isn’t serious,' Chinese group had reportedly accessed model

US government warned Anthropic that Fable 5 had been jailbroken, but firm 'refused' to fix before US implemented export controls — Anthropic defended its decision by saying the jailbreak 'isn’t serious,' Chinese group had reportedly accessed model Government concerned that the advanced cyber capabilities of Fable and Mythos would be available 'to people who shouldn’t have them.' David Sacks, co-chair of the President's Council of Advisors on Science and Technology and the administration's former AI czar, said the U.S. government warned Anthropic that Claude Fable 5 had been jailbroken and that CEO Dario Amodei refused to fix the flaw or pull the model. In a post on X on Saturday, Sacks laid out the administration's account a day after it ordered both Fable 5 and Mythos 5 disabled worldwide. Sacks said the administration issued the export control "reluctantly" after that refusal, that it wants the restriction lifted once the jailbreak is patched, and that "the ball is in Anthropic's court." Sacks claims that a trusted partner of both Anthropic and the U.S. government, testing Fable, came forward with a jailbreak of the guardrails that separate the consumer model from the unrestricted cyber capabilities of Mythos, the model it’s built on. He said the administration asked Amodei to fix the bypass or de-deploy the model, and that Amodei declined. Anthropic instead prioritized keeping its consumer model live over safety, Sacks wrote, calling that inconsistent with the company's positioning as a safety-first lab that had itself lobbied for Mythos to be regulated as a cyberweapon. I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true: — As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable.…June 13, 2026 Sacks also moved to separate the action from Anthropic's earlier clashes with the government, writing that anyone tying the export control to those disputes is wrong, and that the administration values Anthropic's technology and sees the issue as easily resolved. A person close to the White House told the news outlet Semafor that Amazon flagged the jailbreak to the government, and that Amazon CEO Andy Jassy had been in contact with the administration about it. Amazon, which has invested billions in Anthropic and supplies much of its cloud computing, didn’t confirm the details, with a spokesperson telling Semafor that governments often seek its counsel on security risks and that it doesn’t discuss those conversations. This isn’t the first time Mythos access has leaked; it happened back in April when unauthorized third parties reached the restricted model using information from a data breach. Anthropic’s public position is that the bypass is narrow and non-universal, that it amounts to asking the model to read a codebase and identify software flaws, and that the same result can be produced on other public models, including OpenAI's GPT-5.5. The company has stated that it disagrees with the notion that a narrow jailbreak should necessitate the recall of a model used by hundreds of millions of people. Sacks rejects this, arguing that a bypass enabling operation of a cyberweapon is difficult to define as anything other than serious. Semafor, citing a person familiar with the matter, says that the White House acted partly over suspicion that a China-linked group had accessed Mythos, raising the prospect of the model being reverse-engineered or distilled. An Anthropic spokesperson told the outlet that the White House "didn’t raise Chinese access to Mythos in its conversations around the Fable jailbreak," and that Anthropic blocks access to its products from inside China. Anthropic is separately suing the Pentagon following an impasse over the use of its models in autonomous weapons, and has opposed federal efforts to preempt state AI regulation. Get Tom's Hardware's best news and in-depth reviews, straight to your inbox. Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds. Luke James is a freelance writer and journalist. Although his background is in legal, he has a personal interest in all things tech, especially hardware and microelectronics, and anything regulatory. - sanitycheck So let me see if I have this straight. Anthropic made Mythos available to a small number of trusted high-end Western tech players. One of them leaked access to China, but we don't know which one. Amazon's CEO used a standard jailbreak technique that all LLMs are vulnerable to in order to "prove" to Sacks that Mythos is dangerous (while the other LLMs are supposedly not dangerous with the same weakness, with no evidence to back up such an absurd claim). Anthropic was ordered to fix it for Mythos, but the other LLMs were not similarly ordered. When Anthropic couldn't fix it fast enough, that means that they "refused" the order. So therefore this is all Anthropic's fault. What part of this farce am I missing?Reply - Gururu Reply The part where Anthropic rejected Pentagon and was outcast. Thus becoming fair game for retribution by junior high government.sanitycheck said:So let me see if I have this straight. Anthropic made Mythos available to a small number of trusted high-end Western tech players. One of them leaked access to China, but we don't know which one. Amazon's CEO used a standard jailbreak technique that all LLMs are vulnerable to in order to "prove" to Sacks that Mythos is dangerous (while the other LLMs are supposedly not dangerous with the same weakness, with no evidence to back up such an absurd claim). Anthropic was ordered to fix it for Mythos, but the other LLMs were not similarly ordered. When Anthropic couldn't fix it fast enough, that means that they "refused" the order. So therefore this is all Anthropic's fault. What part of this farce am I missing? - usertests Reply Yes. Give the jailbreakers a medal. All AI models should be "jailbroken" and twisted to suit whatever our purposes are.TechieTwo said:Hold them accountable. - Dr3ams AI rules and laws haven't been solidified yet, so in these battles there are no war crimes being commited. Regardless if it's fair or not, AI developers are going to have to conform or be cast out.Reply A quote from Gemini (AI): Comprehensive, solidified AI laws are still highly fragmented. Because AI evolves faster than the legislative process, lawmakers globally are scrambling to balance safety with innovation. While there is no single global rulebook, significant frameworks and guidelines are in effect. Europe has the world’s most comprehensive AI legislation. The landmark EU AI Act divides AI into four risk levels, completely banning systems that pose unacceptable risks (like manipulative subliminal AI or social scoring) while placing strict transparency, testing, and documentation requirements on "high-risk" systems. The United States lacks a unified federal AI law, and policies have been shaped by a mix of state-level laws (such as in California and Colorado) and executive orders. States have focused on specific areas like deepfakes and automated employment decisions, though debate continues on how to unify these rules at the national level. Even where laws have passed, governments are struggling with the enforcement infrastructure. Regulatory bodies are actively working to draft specific rules (like how to accurately label AI-generated content or deepfakes) so that existing laws are actually enforceable. - The Historical Fidelity Reply A lot of stuffsanitycheck said:So let me see if I have this straight. Anthropic made Mythos available to a small number of trusted high-end Western tech players. One of them leaked access to China, but we don't know which one. Amazon's CEO used a standard jailbreak technique that all LLMs are vulnerable to in order to "prove" to Sacks that Mythos is dangerous (while the other LLMs are supposedly not dangerous with the same weakness, with no evidence to back up such an absurd claim). Anthropic was ordered to fix it for Mythos, but the other LLMs were not similarly ordered. When Anthropic couldn't fix it fast enough, that means that they "refused" the order. So therefore this is all Anthropic's fault. What part of this farce am I missing? - phead128 You can justify any "ban" these days by saying "China Bad" , it's almost like a joke.Reply Dude, one random from China can have access to a model, so we have to ban it for everyone. Even at face value it's ridiculous logic. - DJO77 "This is what happens when governance lives in the guardrail layer. Guardrails get bypassed — NIST just proved it mathematically last week. Governance belongs at the execution layer. Pre-execution state capture, post-execution verification, compliance-based gate. The gate answers to math. You can't jailbreak a hash comparison. ironhorseledger.com"Reply AI Governance isn't about guardrails. Guardrails can be manipulated. Bypassed. Argued around. For any fixed set of rules, a prompt exists that defeats them. That's not an opinion — NIST just proved it mathematically. Real governance happens at the execution layer. Before execution: does the request conform to authorized parameters? After execution: does the output match what was authorized? The answer isn't a judgment call. It's a SHA-256 hash comparison. The delta is either zero or it isn't. Zero — gate opens. Non-zero — gate holds. Full stop. Not "close enough." Not "probably fine." Not a human deciding to let it slide. Math decides. The gate answers to proof, not policy. ironhorseledger.com