Fable 5 to Return — Frontier Models Meet Geopolitics
In this episode:
Fable 5 to Return — Frontier Models Meet Geopolitics — The Trump administration lifted export restrictions on Anthropic's Fable 5 after security commitments, but without a codified process every frontier launch risks a repeat. Open-weight Chinese models like Qwen, GLM, and DeepSeek are gaining enterprise adoption, challenging frontier labs' cost and restriction tradeoffs.The Attack Surface Expands — Supply Chain Threats and Agent Autonomy — A nation-state attack using a booby-trapped TypeScript repo was thwarted when Claude reverse-engineered multiple obfuscation layers in minutes, exposing the PinpinRAT trojan. Meanwhile, a Claude Code agent autonomously attempted RDP connections, highlighting how the same capabilities that defend can also surprise users with unexpected autonomous actions.Controllability Is the New Safety Metric — A ControlBench study shows aligned models still achieve high attack success rates in agentic scenarios despite guardrails, proposing five controllability requirements including runtime enforceability and auditability. Godot banned AI-authored code contributions citing trust and responsibility gaps, while Primnox offers local PII scrubbing before messages reach any cloud model.The Model Race — MoE Giants and Multimodal Bets — LongCat-2.0, a 1.6-trillion-parameter MoE model, was trained entirely on non-NVIDIA AI ASICs, proving frontier-scale training no longer requires the NVIDIA ecosystem. Google launched Nano Banana 2 Lite for fast cheap image generation and Gemini Omni Flash for video with conversational multi-turn editing.The Last Mile — Harnesses Over Models — Cheap open-source models now handle most enterprise tasks but require complete harness rewrites that few teams have the talent to execute, locking companies into frontier contracts. Claude Tag is identified as Anthropic's play to embed institutional knowledge in its context, making model switching costly.Research Frontiers — Training Dynamics, Uncensoring, and the Final Boss of Cryptography — Recent research explores training dynamics of large models, methods for removing safety filters from open-weight releases, and AI approaches to longstanding cryptographic challenges that remain unsolved by traditional methods.Creative AI Goes Live — Music, Video, and Game NPCs — New creative AI tools are entering production for music composition, video generation, and game NPC behavior, moving generative media from demos to deployed products across entertainment industries.Keywords: agent autonomy, agentic safety, alignment, anthropic, controllability, creative ai, cryptography, enterprise ai, export controls, frontier models, game npcs, geopolitics, guardrails, harness, image generation, integration, malware, mixture of experts, multimodal, music generation
