Anthropic Publishes Its Full Jailbreak Evaluation Framework for Fable 5
π¬ RESEARCH
Anthropic Publishes Its Full Jailbreak Evaluation Framework for Fable 5
Anthropic just open-sourced the red-team methodology behind Fable 5's cyber safeguards β automated attack generation, graded severity taxonomies, regression suites, the whole playbook. This is the first time a frontier lab has published its jailbreak evaluation framework at this level of detail, and it's a deliberate move: if everyone can see how you test, everyone can hold you accountable. The bet is that transparency accelerates defense faster than it arms attackers. If you run safety evals, read the framework and compare it to yours. Read more β
High-Severity CVEs Spiked Around the Mythos Preview Launch Window. Epoch AI's analysis shows a statistically significant spike in newly reported high-severity CVEs timed around the Claude Mythos Preview release. Whether this is correlation or causation, it's the first empirical data point linking frontier model releases to vulnerability discovery rates β and it's going to shape every future model-release safety debate. (18 likes | 4 RTs) Read more β
Dispersion Loss Stops Small Language Models from Collapsing Their Embeddings. If you're training sub-3B models, you've hit this: the embedding space collapses, and quality tanks. Dispersion loss is a novel training technique that counteracts this condensation problem β effectively a free performance boost for SLMs with no architecture changes required. Worth adding to your next fine-tuning run. (15 likes | 3 RTs) Read more β
π§ LAUNCH
HuggingFace Transformers Ships Native KimiK 2.5β2.7 Support
Transformers v5.13.0 adds first-class architecture support for the KimiK 2.5β2.7 family β Moonshot's open-source multimodal agentic models. If you've been waiting for proper HuggingFace integration before testing these, the barrier just dropped to a single pipeline() call. The multimodal and agentic capabilities make them worth benchmarking against whatever you're currently running. Read more β
Diffusers Brings NVIDIA Cosmos 3 Into the Standard Pipeline. Diffusers v0.39.0 ships support for Cosmos 3 β NVIDIA's unified world foundation model built on Mixture-of-Transformers. One model now handles world generation, physical reasoning, and action generation through the familiar diffusers API. If you're prototyping anything in physical AI or simulation, this is the fastest path to testing it. Read more β
π‘ INSIGHT
DeepMind and A24 Launch the First Frontier-AI-Meets-Film-Studio Research Partnership
Google DeepMind and indie-film powerhouse A24 are partnering to research AI's role in professional filmmaking β the first formal collaboration between a frontier AI lab and a major creative studio. This isn't about replacing filmmakers; it's about figuring out which creative workflows actually benefit from AI tooling and which don't. Expect this deal to set the template for how every other studio approaches the question. Read more β
AIEWF Wraps With a Fight Over Agentic Loops and an H2 Roadmap. The AI Engineer World's Fair closed with a heated debate on when agentic loops help versus when they're just expensive retry machines. The closing keynotes laid out a state-of-AI-engineering report that's genuinely useful for H2 planning β read the summary before your next team offsite. Read more β
Vercel's Chief of Software: Skills, Sandboxes, and Agent-Readable Websites Are the New Primitives. Andrew Qu explains how Vercel built their agent framework "eve" and argues the web needs to be restructured for AI consumers, not just human ones. The concrete claim: agent-readable websites will matter as much as mobile-responsive design did a decade ago. Whether or not you buy Vercel's specific bet, the primitives he outlines β skills, sandboxes, structured tool interfaces β are where frontend architecture is heading. Read more β
Simon Willison Maps the Gaps Between Open-Source and Closed AI. Willison's latest gap map is a clear-eyed inventory of where open-source models still fall short of closed providers β and where the contributor opportunity is widest. If you're betting on open models in production, check whether your stack covers the gaps he flags before you ship. Read more β
π§ TOOL
Claude Code Renames 'Default' to 'Manual' and Stops Auto-Continuing Dialogs. Claude Code v2.1.200 renames the "default" permission mode to "Manual" across all surfaces and stops AskUserQuestion dialogs from auto-continuing β a meaningful change if you run Claude Code in semi-autonomous mode. v2.1.201 drops mid-conversation system-role reminders for Sonnet 5 sessions. Review your permission-mode config after updating. Read more β
llama.cpp Fuses 288-Expert MoE Routing Into a Single CUDA Kernel. Three llama.cpp releases in two days. The headline from b9866: CUDA topk-MoE fusion for 288-expert models like Step-3.7-Flash, cutting routing overhead from ~330 extra kernel launches to one fused call at batch size 1. b9867 adds spec-draft-p-min support in DFlash for better speculative decoding. Update if you run large MoE models locally. Read more β
mcpsnoop: Wireshark for Model Context Protocol Debugging. Debugging MCP tool calls has been a black box β mcpsnoop fixes that by sitting as a transparent proxy between client and server, giving you a live TUI of every message. Purpose-built for the Model Context Protocol, not a generic TCP sniffer. Essential if you're building or debugging MCP integrations. (44 likes | 13 RTs) Read more β
π TECHNIQUE
A Practical Guide to Running Frontier-Class Open Models Locally. A well-received guide (254 HN points) that walks through hardware selection, quantization tradeoffs, and serving configurations for running current SOTA open models on your own metal. Covers the decisions most guides skip β which quantization format actually matters for your use case, when to use speculative decoding, and where the cost-performance cliff lives. Bookmark this if you're evaluating on-prem inference for privacy or latency. (254 likes | 120 RTs) Read more β
π MODEL LITERACY
Jailbreak Evaluation Frameworks: Anthropic just published theirs in full, so this is a good day to understand what they actually are. A jailbreak evaluation framework is how labs systematically probe their own models for safety failures β think automated attack generation (thousands of adversarial prompts crafted to bypass safeguards), graded severity taxonomies (not all jailbreaks are equal β leaking a recipe is different from generating exploit code), and regression suites that ensure each new model version doesn't reintroduce old vulnerabilities. Understanding this machinery explains why "jailbreak-proof" is a moving target, not a finish line: every patch creates new attack surfaces, and every new capability opens vectors that didn't exist before. The framework has to evolve as fast as the model does.
β‘ QUICK LINKS
- GLM-5.2 on AMD MI355X: Hits 2,626 tok/s per node at over 2x lower cost than Blackwell β AMD's accelerators are getting real. (32 likes | 9 RTs) Link
- Willison on Fable 5's Ethical Judgment: Independent analysis of how Fable handles ambiguous moral scenarios β read alongside Anthropic's own safeguards post. Link
- Vercel AI SDK v7.0.12β7.0.13: Fixes tool-result ordering in agent chains and a whitespace bug in extractJsonMiddleware. Link
- Talos: Earn GPU payouts by serving open-model inference over WebSocket β decentralized inference for idle hardware. (307 likes | 12 RTs) Link
- deptrust: CLI that scans dependencies for vulnerabilities before AI agents pull them in β addresses a real attack surface in agentic coding. (7 likes) Link
π― PICK OF THE DAY
Anthropic open-sources its jailbreak evaluation framework β and dares every other lab to follow. Publishing your full red-team methodology is a move that sounds reckless until you think about the game theory. If Anthropic keeps its attack playbook private, it stays one lab's advantage. If it publishes, every security researcher, every competitor, and every academic can run the same probes against every model β including Anthropic's own. That's the bet: transparency accelerates defense faster than it arms attackers. The framework includes automated attack generation, severity taxonomies, and regression suites β the kind of infrastructure that takes years to build and that most labs guard like crown jewels. By publishing, Anthropic forces a choice on every competitor: match the transparency or explain to regulators and customers why you won't. This is safety-as-competitive-strategy, and it's the smartest move in the AI trust game right now. Read more β
Until next time βοΈ