Claude Fable 5

Claude Fable 5 and Mythos 5 are new Mythos-class models delivering SOTA performance in software engineering, vision, and scientific research. Fable 5 features conservative safety classifiers that fallback to Opus 4.8 for high-risk cybersecurity or biological queries, while Mythos 5 provides unrestricted access for trusted partners via Project Glasswing. These models demonstrate significant gains in long-context autonomy, novel hypothesis generation, and token efficiency, priced at $10/M input and $50/M output tokens. A mandatory 30-day data retention policy is now required for all Mythos-class traffic to facilitate safety monitoring and jailbreak detection.

Read more

CEOs who think AI replaces their employees are just bad CEOs

CEOs often overestimate LLM capabilities because they are disconnected from the "last mile" of production work, mistaking "happy path" prototypes for scalable, compliant solutions. While LLMs enhance productivity when used as assistive tools, they do not replace the human expertise required for security, legal review, and complex systems integration. Effective implementation requires understanding AI limitations rather than using it as a pretext for headcount reduction or forcing counterproductive metrics like token leaderboards.

Read more

Microsoft's open source tools were hacked to steal passwords of AI developers

Microsoft has disabled approximately 70 open-source GitHub repositories following a supply chain attack that injected credential-stealing malware into Azure and AI development tools, including CLI interfaces for Claude and Gemini. The malware targets sensitive credentials when developers integrate the compromised tools into AI coding workflows. This incident marks the second breach of Microsoft’s open-source ecosystem in recent weeks, following a previous compromise of the Durable Task project.

Read more

Cleaning up after AI rockstar developers

The "rockstar developer" archetype, known for high-velocity but unmaintainable and overly clever code, is being replicated at scale through LLM-generated "vibe coding." This process creates fragmented technical debt across disconnected contexts, leading to systems so complex they require AI just to navigate. To maintain software longevity, engineers should use LLMs for granular tasks under strict human oversight, prioritizing architectural simplicity and readability over raw generation speed.

Read more

German ruling declares Google liable for false answers in AI Overviews

The Regional Court of Munich ruled that Google is directly liable for its AI search overviews, classifying them as original content rather than traditional search results. The court found that because the LLM synthesizes, rewrites, and generates independent claims not present in linked sources, Google acts as a direct infringer rather than a mere intermediary. This precedent strips LLM providers of traditional search engine liability shields, holding them responsible for hallucinations and defamatory outputs generated by their algorithms.

Read more

AutoMegaKernel: Compiling a LLM into a single CUDA kernel

AutoMegaKernel (AMK) compiles Llama-family models into a single persistent cooperative CUDA kernel, utilizing a schedule-IR validator to statically ensure deadlock and race freedom. The system supports multiple NVIDIA architectures (sm_80/90/120) and employs an agent-driven search loop to optimize performance across various model architectures. While AMK's W8A16 kernels outperform cuBLAS bf16 on inference-class GPUs like the L40S and RTX 5090 at batch-1 decode, it currently trails on high-bandwidth training-class hardware due to cross-SM synchronization bottlenecks.

ArXiv

Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories

Analysis of 20,000 LLM-generated stories reveals significant low variability, with 11 specific tokens appearing in 88.3% of outputs across multiple models. These patterns are traced back to small preference datasets used during alignment rather than pre-training data or general literature. The findings demonstrate that alignment algorithms can disproportionately narrow model output diversity based on limited fine-tuning data.

ArXiv

Configuring Agentic AI Coding Tools: An Exploratory Study

An empirical study of 2,853 GitHub repositories identifies eight configuration mechanisms for agentic AI coding tools, ranging from static context to executable subagents. Findings show that static Context Files, particularly the interoperable AGENTS.md standard, are the primary configuration method, while advanced features like Skills remain underutilized and largely non-executable. Adoption patterns vary by tool, with Claude Code demonstrating the most diverse configuration usage among the analyzed platforms.

ArXiv

Steganography Without Modification: Hidden Communication via LLM Seeds

Researchers discovered a steganographic channel within LLM inference stacks that exploits the seed-dependent sequence of token-level probability intervals produced by PRNGs during deterministic decoding, requiring no modification to model weights or output distributions. A sender encodes a secret message in the PRNG seed, which a receiver can recover by reconstructing these intervals from the generated text, even in unknown-prompt settings via approximate reconstruction. Experiments demonstrate high accuracy for 32-bit seed recovery within hundreds of tokens and seconds, enabling covert data transmission and indicating that prompt ignorance is not a valid security assumption for LLM outputs.

ArXiv

Is Grep All You Need? How Agent Harnesses Reshape Agentic Search

This empirical study evaluates the impact of retrieval strategies and tool-calling paradigms on LLM agent performance using harnesses like Chronos and provider-native CLIs. Findings demonstrate that grep-based retrieval generally outperforms vector search in accuracy, though results remain highly sensitive to agent architecture and the presence of distracting context. The research highlights how tool output presentation and noise in conversation history significantly influence agentic RAG workflows.

ArXiv

Learn from 30 historical figures, open source, nonprofit, self-hosted

Agora Cosmica is an open-source, AGPL-3.0 licensed platform that enables interactive dialogues with historical figures using a privacy-first architecture. It leverages Qwen3 235B for inference, supports BYOK with local AES-256-GCM encryption, and features a self-hosted audio stack utilizing Kokoro, Qwen3-TTS, and Faster-Whisper. The system is designed for local deployment via Docker and supports integration with OpenAI-compatible endpoints like Ollama, vLLM, or LM Studio.

Repo

Nucleus – A security-hardened, Nix-native container runtime

Nucleus is an extremely lightweight, security-hardened, declarative Linux container runtime designed for AI agent workloads and production services. It provides fast-startup (12ms cold start) ephemeral sandboxes for AI agents, utilizing Linux kernel primitives like namespaces, cgroups, seccomp, and Landlock for zero-overhead isolation. Nucleus features "Agent mode" for quick, context-prepopulated execution and "Strict agent mode" for fail-closed isolation. Deeply integrated with Nix, it enables reproducible, declarative root filesystems and security policies, offering auditable and stable runtime environments for untrusted or ephemeral workloads, distinct from traditional Docker-style image management.

Repo

Claw Patrol, a security firewall for agents

Claw Patrol is a security firewall for agents, intercepting their wire-level traffic to production systems. It enforces actions via HCL rules, leveraging CEL expressions over protocol-specific facts (e.g., SQL verbs, Kubernetes resources, HTTP details). This enables granular control, such as blocking destructive commands or requiring human approval for sensitive operations, and can be deployed as a central gateway, a host-wide tunnel, or a per-process wrapper.

Repo

AutoGPU – AI designs a real 7nm GPU, from Verilog to GDSII

AutoGPU showcases AI agents autonomously designing a 7nm GPU, managing the entire flow from Verilog RTL to a GDSII layout through iterative synthesis, place-and-route, and verification. Agents collaborate, open GitHub issues, and review PRs, with humans programming their behavior and workflows by editing Markdown documentation. The system has produced a hardened fp8 matmul accelerator (32x32 systolic array) with a clean GDSII layout, a self-written sign-off toolchain, and web viewers, with full chip integration actively being developed.

Repo

Deep Memory – Vocabulary-driven graph memory for AI agents

@utaba/deep-memory is a vocabulary-driven graph memory library for AI agents, providing structured, persistent knowledge graphs. It addresses the cold-start problem by offering a governed schema (vocabulary) of entity types, relationships, and properties, enabling agents to create consistent data and traverse graphs efficiently with optimized token usage. The system exposes 20+ tools via an MCP server and includes an optional AI-driven indexing pipeline to transform documents into structured KGs, supporting pluggable storage, search, and embedding providers.

Repo