Claude Mythos breaks NIST post-quantum cryptography in 60 hours

Anthropic's Mythos breaks NIST post-quantum cryptography standard in 60 hours

Mythos identified a mathematical weakness in HAWK, a NIST post-quantum digital-signature standard that had survived two years of expert review, forcing the HAWK team to withdraw from competition the next day. The model also improved attacks on 7-round AES-128 by 200–800×, raising immediate questions about cryptographic security as more capable AI systems become available.

Google acquires Spirit Airlines bankruptcy data for $10M to train AI models

Google purchased 100 million emails, 500 million Teams chats, 7.5 billion passenger transaction records, 7.2 billion competitor flight prices, and 30 million lines of code from Spirit's bankruptcy estate. The deal signals an emerging market for salvaging corporate datasets as AI systems become better at extracting value from raw operational records previously worthless outside their original context.

Stripe acquires OpenRouter for $7.5B, positioning itself as AI token marketplace infrastructure

Stripe is paying $7.5B ($1.5B to founders, $6B to investors) for the model-agnostic API router, which CEO Patrick Collison frames as critical infrastructure for managing "token flows" alongside revenue flows as businesses adopt multiple AI providers. The acquisition signals major consolidation in the AI inference layer and Stripe's strategic pivot toward becoming core infrastructure for AI cost management and multi-model orchestration.

Post-training scaling and agent harness architecture emerge as primary competitive layers, displacing parameter count

Z.ai's GLM-5.3 demonstrates the shift decisively: using the same base model as GLM-5.2, it achieves substantial capability gains through one month of RL training on long-horizon environments (multi-day engineering tasks with access to compute clusters, codebases, and live experiment results). GLM-5.3 improves from 4.6 to 28.3 on Terminal Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 24.4 to 54.4 on ExploitBench, while also discovering 2,436 real-world vulnerabilities across 269 projects—many unnoticed for decades. Prof Jie Tang argues parameter count is now a misleading metric; the scaling frontier has shifted decisively toward training recipe quality, data composition, and post-training methodology. This structural shift is reshaping competitive positioning across the stack. Open-weight models (Ornith-1.5, GLM-5.3, Kimi K3) now dominate agentic benchmarks, with managed harnesses and token routing becoming the primary differentiation layer. TrueForge matches Claude Managed Agents on Opus 4.8 while using ~30% fewer tokens; Microsoft's Agent Lightning improved Qwen 3.5-9B from 41.8% to 56.4% on SWE-Bench Verified with only 6K training examples. DeepSeek Harness's plugin-first design and TrueForge's vendor-neutral, self-hostable stack signal that the session/environment/memory/tools layer is now a major source of both differentiation and cost savings—mirroring broader infrastructure consolidation where RL-through-the-harness and retrieval system optimizations (Qdrant's filterable HNSW, Sentence Transformers v6.0 multi-vector shift) are becoming standard production recipes.

Discovering Cryptographic Weaknesses: Claude Mythos Breaks HAWK Post-Quantum Signature Scheme and Improves AES Attacks

Anthropic published a research report on Claude Mythos Preview's autonomous cryptanalytic capabilities. Main takeaways:

• Mythos discovered an improved attack on HAWK, a NIST post-quantum signature candidate, reducing effective keysize by 50% in 60 hours of semi-autonomous work (~$100k API cost), forcing key size doubling to maintain security.
• Mythos improved attacks on 7-round AES by 200-800x speed improvement via meet-in-the-middle techniques, though attack remains impractical and does not affect full 10-round AES.
• Results demonstrate frontier AI models can autonomously discover mathematical flaws in cryptographic algorithms before deployment, with implications for cryptographic standardization processes and security review workflows.
• Multi-agent collaborative harness enabled discovery; human operator provided project management only, not cryptographic expertise.

Transformers Math 101: Training Rules of Thumb with Quentin Anthony

Latent Space published a podcast interview with Quentin Anthony from EleutherAI on transformer training mathematics and scaling laws. Main takeaways:

• Core compute equation: C = τ × T = 6 × P × D, where P is parameters, D is dataset tokens, τ is hardware throughput (FLOPs/GPU × GPU count), and T is training time; enables cost estimation for model training projects
• Chinchilla scaling law (C = 120P²) assumes compute fungibility across time/parallelism; breaks down with communication overhead at scale, requiring adjustments for real distributed training
• Practical memory budgeting requires accounting for model weights, optimizer states (Adam uses 2x parameter memory), gradients, and activation recomputation; ZeRO sharding and 3D parallelism (data/tensor/pipeline) are essential for scaling beyond single GPU
• EleutherAI's approach prioritizes sharing tacit knowledge from large-scale training (GPT-NeoX, others) as open technical documentation rather than papers, filling gap between theory and production systems engineering

U.S. Prepares to Force Countries to Choose Sides in AI Race with China, Threatening Exclusion from Pax Silica

Reuters reported on a leaked State Department draft letter signaling the U.S. will demand countries choose between joining the U.S.-led Pax Silica AI coalition or China's competing World Artificial Intelligence Cooperation Organization. Main takeaways:

• U.S. explicitly threatening exclusion from Pax Silica for countries that also join China's AI framework, framing it as binary choice rather than non-binding cooperation
• Kazakhstan identified as only country currently in both initiatives, triggering U.S. alarm and prompting the ultimatum letter
• Pax Silica membership provides access to shared AI investment opportunities and critical mineral supply chains; China's framework promotes open-weight models as alternative to U.S. proprietary systems
• U.S. strategy aims to starve China of resources and allies amid Chinese open-weight models making rapid gains against OpenAI and Anthropic systems
• Reflects escalating geopolitical bifurcation of AI infrastructure, supply chains, and model ecosystems along U.S.-China lines