Washington Pulls Anthropic's Most Powerful Models as a Trillion-Dollar IPO Summer Begins

Policy & Safety

• On Friday at 5:21 pm ET, the US Commerce Department issued an export control directive ordering Anthropic to block all foreign-national access to Claude Fable 5 and Mythos 5 — inside and outside the US, including the company's own foreign employees — citing national security concerns. To comply, Anthropic abruptly disabled both models for all customers worldwide while leaving its other models available, calling the move a 'misunderstanding' it hopes to reverse. The underlying concern, per an Axios source, is a reported jailbreak that bypasses classifier-based safeguards on cybersecurity, chemistry, and biology prompts; the administration wants a pause of a few weeks to let the 'national security apparatus' harden. Anthropic says it has only received verbal evidence of a 'narrow, non-universal jailbreak' — essentially asking Fable 5 to review a codebase for software flaws — and argues that capability is already widely available in models like OpenAI's GPT-5.5. The company warned that recalling a model deployed to hundreds of millions over a narrow jailbreak, if applied industry-wide, 'would essentially halt all new model deployments for all frontier model providers.' ^{[4][12][14][15]}
• A Ukrainian drone maker revealed that fully autonomous AI drones killed Russian soldiers in a one-time battlefield test roughly two years ago, marking a milestone in autonomous lethal weapons. Aero Center CEO Alexander Kokhanovskyy described quadcopters preprogrammed to fly to a front-line area and activate an AI-powered 'Terminator mode' to seek and attack any target; human-piloted drones later found 'a couple' of dead soldiers. Ukrainian officials stressed the government bans AI in the final stage of target interception and that pilots use semi-autonomous systems with humans in the loop. The context matters because AI-driven navigation has lifted Ukrainian strike success rates from 10–20% to 70–80%, and Ukraine now launches over 5,000 drone strikes monthly beyond 20 km, showing how autonomy is quietly spreading even where full autonomy remains banned. ^[21]
• Google filed a lawsuit to dismantle Outsider Enterprise, an alleged Chinese cybercrime network that used AI — including Google's own Gemini — to automate phishing scams against 'hundreds of thousands of victims.' The group's 'phishing-for-dummies' software, sold for $88 per week or $200 per month, offered nearly 300 templates to clone real sites and instructions to 'weaponize AI-generated code,' deploying 9,000 fake websites, 1 million fraudulent domains, and 2.5 million scam texts to Android users. The FBI, working with Google and Lumen's Black Lotus Labs, seized domains and Shopify storefronts, and said the platform has enabled theft of an estimated 3.87 million credit cards and $1.9 billion in losses since July 2023. Google, which says its tools intercept over 10 billion scam messages a month, is seeking damages and an injunction, signaling labs will increasingly police criminal abuse of their own models. ^[18][28]