Anthropic's Mythos Model Broke Into Almost All US Classified Systems in Hours
"This tool broke into almost all of our classified systems, not in weeks but in hours. Mythos identified the holes with exploitation outside the scope of the exercise."
About this episode
In an emergency weekend recording of Moonshots, hosts Peter Diamandis, Alex Guzey, Dave Blunden, and Imad Mustaq addressed explosive developments in AI regulation and geopolitics. The most significant story: for the first time in US history, the executive branch has placed national security holds on commercial AI products, requiring government approval for customer-by-customer release of OpenAI's GPT-5.6 and Anthropic's Mythos models. The trigger was Project Glasswing, where Mythos broke into almost all classified US systems within hours during red team exercises with intelligence agencies. Senator Mark Warner's revelation about the cyber capabilities prompted immediate restrictions on foreign national access. The episode explored whether this regulatory regime will crush domestic innovation while China's open-weight models rapidly achieve parity, with analysis showing Chinese capabilities could match Western frontier labs by Christmas 2025. Anthropic accused Alibaba of the largest AI theft in history, alleging 28.8 million fraudulent queries across 25,000 fake accounts to distill Claude's capabilities. The hosts debated whether US export controls arrived too late, given that existing models with proper harnesses already exceed restricted frontier capabilities. Beyond AI regulation, the episode covered Elon Musk's announcement that Neuralink will attempt the first human brain-to-brain telepathic communication later in 2025, Trump's $2 billion quantum computing executive order targeting companies like IBM and PsiQuantum, ByteDance's SeaDance 2.5 video generation model producing 30-second 4K videos, and Eli Lilly's $6.3 billion acquisition of orexin-targeting drug technology that could enable humans to function on four hours of sleep. The consensus: we are in a hard takeoff toward the singularity, with every week bringing more developments than the last.
Key takeaways
- US government now controls frontier AI model releases customer-by-customer after Anthropic's Mythos cracked classified systems in hours during intelligence agency tests.
- Anthropic accused Alibaba of running 28.8 million fraudulent queries across 25,000 fake accounts to steal Claude's capabilities through distillation in largest AI theft claim ever.
- Chinese open-weight models on trajectory to match Western frontier capabilities by Christmas 2025, with GLM 5.2 already outperforming GPT-5.5 on hardest benchmarks when properly harnessed.
- OpenAI delayed IPO and GPT-5.6 release amid regulatory pressure, valuing staying private over $1 trillion public valuation to focus on hard takeoff rather than SEC compliance.
- Neuralink announced plans for first human brain-to-brain telepathic communication attempt later in 2025, shifting from medical device to superhuman enhancement technology.
- Trump signed executive orders committing $2 billion to quantum computing with IBM receiving $1 billion, establishing quantum protection team amid concerns over Chinese espionage.
- Eli Lilly acquired orexin-targeting drug technology for $6.3 billion that could enable humans to function on four hours of sleep, potentially adding 58 days of productivity per year.