Frontier labs dismiss encrypted reasoning theft, researchers extract it

Johns Hopkins cryptographer Matthew Green reveals that frontier AI APIs return encrypted reasoning blobs that can be replayed across sessions, users, and models: Anthropic, OpenAI, and Google send encrypted chain-of-thought blocks back to API clients, and those blocks carry no binding to the session, user, or model that produced them. ELLIS Institute Tübingen PhD researcher Alexander Panfilov and collaborators extended this finding: by injecting a stronger model's encrypted trace into a weaker sibling and jailbreaking that sibling, they recover the frontier model's hidden reasoning in plaintext without attacking the stronger model directly or triggering its anti-distillation safeguards. Johns Hopkins cryptographer Matthew Green had reported the replay flaw to the labs in May, but they dismissed any security implications. The Stolen Thoughts project site includes live decoded examples and a "guess the model" game.

Company Offering '100% Human-Written, Never AI' Medical Research Is Entirely AI: 404 Media's Emanuel Maiberg investigated Research Gold, a service selling peer-review-ready medical manuscripts and systematic reviews while claiming human-only authorship. The PhD reviewers listed on its site are AI-generated personas that don't exist, other named staff are real people whose identities are used without permission, and every customer touchpoint, phone, email, and chat, is handled by AI. The story lands as journals already struggle with AI-generated papers and compromised peer review.

Signing and verifying AI outputs with revocable trust using BlkSeal: lyfe.​ninja's BlkSeal platform lets developers cryptographically sign AI responses, API outputs, and documents at creation time, then verify origin, detect tampering, and revoke trust if the signing authority is later invalidated. The talk walks through how revocable signatures work in practice, addressing a real gap as AI-generated content becomes harder to distinguish from human-created content.

Fully autonomous robots are much closer than you think, Sergey Levine on the Dwarkesh Podcast: UC Berkeley professor and Physical Intelligence co-founder Sergey Levine argues a "self-improvement flywheel" is beginning for general-purpose robots, with his median estimate for fully autonomous household robots landing at 2030. The conversation covers how vision-language-action (VLA) models work, why robotics will scale faster than self-driving cars, the role of simulation in training, and whether China's hardware lead gives it a structural advantage.

CMU's LLM Inference Algorithms lecture series is now on YouTube: Carnegie Mellon University Associate Professor Graham Neubig's course CS 11-664/763 covers the full stack of inference algorithms for language models, from transformer basics and search algorithms to modern efficiency techniques. The course site has slides and assignments to go alongside the videos.

Why LLMs fail the "car wash" reasoning test: The question "The car wash is 100 meters away, should I walk or drive?" seems trivial, but Claude, GPT-4, and Gemini all answer "walk." The correct answer is drive, because the car itself must get to the car wash. The video demonstrates the failure across models and explores why next-token-prediction systems miss the implicit object-level constraint. A February 2026 arXiv paper did a variable isolation study on the same problem, finding that prompt architecture (specifically a structured STAR format) largely resolves the failure.

Google's deep-dive video maps the path from basic RAG to knowledge graphs, persistent memory, and multimodal agents: The 90-minute session walks through a nine-step production roadmap: turning raw data into a knowledge graph, adding semantic and hybrid search with Graph RAG, extracting graph context from images, text, and video, then orchestrating specialized agents with Google's Agent Development Kit (ADK). A companion Google Cloud blog post covers building GraphRAG applications with Spanner Graph and LangChain for teams ready to go hands-on.

Graph Engineering with Claude Code: a live terminal walkthrough of agentic graph workflows: This tutorial video walks through building a full Claude Code setup and multi-agent graph workflow from a blank terminal, covering the shift from self-improving loops to structured agentic graphs. Note that the quoted Anthropic engineer framing is disputed by commenters who call the video a basic Claude Code and VS Code integration walkthrough, not a statement from Anthropic. The companion written guide on aibuilderclub covers the same 14-step graph-architecture roadmap in text form.

How Claude marks AI-generated content with invisible text watermarks: Anthropic's new policy embeds machine-readable watermarks in all text generated by Claude models launched on or after August 2, 2026, and adds signed provenance metadata to supported files. GPTZero co-founder and CTO Alex Cui explains that almost every frontier lab watermarking scheme follows the KGW method: before each token is generated, tokens are split into "green" and "red" lists using the preceding context as a seed, and the model is softly nudged to prefer green tokens, leaving a statistically detectable signal that is invisible to readers. The signal survives copy-paste but can be degraded by paraphrasing through a model that doesn't use the same scheme.

DoorDash launches Flux, an internal cloud platform for running AI agents at enterprise scale: Flux moves DoorDash's agentic workloads off individual developer laptops and into a centralized cloud runtime with enterprise security guardrails, enabling it to automate 130k engineering tasks in a single month and handle 25k+ AI-assisted code reviews per week. The post explains the architecture for teams looking to scale beyond local agent harnesses.

Hugging Face's speech-to-speech repo now supports a VAD → multimodal LLM → TTS pipeline, dropping the STT step entirely: Hugging Face Multimodal Research Lead Andrés Marafioti updated the open-source speech-to-speech framework to route audio directly into a multimodal LLM, skipping transcription. Beyond cutting latency, eliminating STT preserves tone, hesitation, and emphasis that a transcript would flatten. The tradeoff is that audio-capable MLLMs (such as Gemma 3's small multimodal variants) are heavier than a dedicated Whisper model, and full-duplex support is still a missing piece.

xAI launches Grok Bot in early beta: persistent AI teammates that sign into your apps and do real work: Grok Bot gives each user a named, long-running AI agent that operates a persistent cloud computer, logs into tools on your behalf, and only returns when it needs approval. That removes the need to keep a local machine running for always-on automation. The launch extends Cursor's agent technology beyond coding into sales, support, finance, and other computer-based work, and Cursor product leader John Bai called it one of the biggest projects he has worked on. Cursor engineer Jediah Katz says the product has spread wildly inside his team and he used a bot named Wonton to snag fast-selling Forbidden City tickets during a China trip.

OpenMausBot: an open-source Grok Bot clone that needs no extra subscription: Hours after xAI launched Grok Bot (an "AI teammate" that logs into your tools and completes tasks autonomously) in early beta, developer Milind Soni released OpenMausBot, a fully open-source equivalent. It spins up virtual machines via Ascii, handles computer-use tasks via Cua, and adds plugin integrations through Composio, all running on subscriptions you already hold rather than requiring a new one. Routines and further integrations are planned, and the project is accepting pull requests.

Continuous Autoregressive Language Models (CALM) proposes replacing next-token prediction with next-vector prediction: CALM compresses chunks of K tokens into a single continuous vector using a high-fidelity autoencoder (over 99.9% reconstruction accuracy), so the model generates whole semantic chunks per step instead of one discrete token. This reduces the number of generative steps by a factor of K and sidesteps the sequential bottleneck that limits today's LLMs. Code and pre-trained checkpoints are available on GitHub.

Capy v2 launches as a cloud coding agent claiming top scores on DeepSWE at half the price of rivals: Capy v2 spawns fleets of reinforcement-learning-trained (RLM) agents, each in its own VM with up to 32 vCPU and 128 GB RAM and sub-second boot times. It claims to outperform Claude Code, Codex, Devin, and Cursor on the DeepSWE coding benchmark while being 50% cheaper. Users can bring their own Codex or Grok subscriptions.

Meta open-sources Muse Glimmer, a 30B agentic model designed to run on consumer hardware: Meta's Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows, released under Apache 2.0. Magnitude founder Tom Greenwald mapped out real-world throughput by GPU: an RTX 3090 hits roughly 40 tok/s on a 4-bit quant and nearly matches the RTX 4090 at half the price, the RTX 5090 reaches 75 tok/s, and a DGX Spark manages only 15 tok/s despite supporting 8-bit quant. Apple Silicon lands at 22 tok/s on an M5 Pro and 27 tok/s on an M5 Max. DFlash speculative decoding boosts speeds across all hardware. The bandwidth-bound nature of dense models explains why the older 3090 punches well above its price class.

NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI: Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts open model tuned for long-running agentic workloads, available free on OpenRouter. NeMo Switchyard is a new open-source routing library that intelligently directs agent requests to the most capable model across edge, PC, workstation, data center, and cloud deployments. The model is also available on Hugging Face.

WorldClaw: Agentic 3D Open-World Generation at Scale: Tencent's Hunyuan3D team presents WorldClaw, a coarse-to-fine pipeline that turns a text prompt into a large, explorable 3D world. Planning agents build a structured scene spec, then terrain, assets, and materials are assembled in layers, with render-based agents doing final refinement. The most novel part: an image model handles region composition (something image models do well), then objects are extracted into 3D meshes via techniques like SAM3D and placed in the world. The code isn't yet public, only pre-generated demo worlds are available, and commenters note the approach is largely an LLM wired to a standard procedural content generation system rather than a new model.

OpenSSH 10.5 released, explicitly welcoming AI-assisted security bug reports: The August 11 release notes include an unusual policy statement: the team has been receiving a surge of AI-assisted security reports and says it "very much welcomes" them, especially when paired with human triage and proposed fixes. Because bugs found by AI are sometimes independently rediscovered by outside researchers (a sign adversaries may find them too), the project will now ship more frequent point releases rather than batch fixes until the next planned version. The release itself also drops support for the legacy DSA key type and requires ECC support from the underlying OpenSSL.

H3-metal, Native MiniMax-H3 inference for Apple Silicon: Redis creator Salvatore Sanfilippo built h3.​c, a pure-C Metal-accelerated inference engine for MiniMax's H3 video model on Apple Silicon Macs. On an M5 Max with 128GB unified memory it generates a roughly 9-second clip in minutes rather than the hour-plus needed by GGUF-based ComfyUI pipelines. The model can also run on 64GB machines with small code changes, though throughput will be limited by memory bandwidth (300 GB/s on M-Pro chips vs. 600 GB/s on M-Max), and weight streaming from SSD instead of RAM adds another bottleneck.

Apple Silicon and macOS VMs: 11, 16× Faster LLM Inference with Llama.​cpp: macOS VMs built on Apple's Virtualization.​framework expose only a conservative Metal GPU profile to guest processes, causing llama.​cpp to run slower, older GPU code paths. The Cua team built a process-scoped Metal capability shim that intercepts and upgrades the capability answers for a single guest process, letting llama.​cpp select newer Metal kernels without affecting the rest of the VM. Benchmarks on an M1 Ultra host show 7.2× faster prompt processing and 14.5× faster token generation. The work is open-sourced under the same permissive license as Lume, M1 Pro and M3 Pro results are not yet available.

llama.​cpp gets an official home at llama.​app: The local-inference engine that runs open-weight models on consumer hardware now has a polished landing page with one-line install, docs, and a model browser. GGML, the team behind llama.​cpp, joined Hugging Face in February 2026, so HF now stewards the project. The HN discussion highlights practical tips: llama-server supports multi-model configs via a single ini file, quantized Qwen3 MoE models punch above their weight for coding tasks, and llama.​cpp pulls models directly from Hugging Face.

What I learned by putting GitHub Copilot behind a MitM proxy: Rafael, writing for the Lighthouse AI newsletter, routed VS Code's Copilot traffic through mitmproxy to see exactly what gets sent. Because Copilot runs inside Electron (which uses Chromium's TLS stack), certificate pinning is not an obstacle, making interception straightforward. Key findings include how Copilot assembles its context window from recent edits and session memory via its Chronicle tool, how auto model selection is negotiated at the network layer, and how much telemetry flows beyond what your prompts contain. Readers in the Hacker News discussion note that eBPF uprobes offer an even cleaner bypass for tools that bundle a statically compiled BoringSSL, letting you capture plaintext right before encryption without fighting any proxy.

Developers are questioning whether Claude is worth it as trust in Anthropic erodes: A developer thread argues that Claude Code's momentum is stalling as users hit low rate limits, notice quality regressions in newer models, and feel condescended to by Anthropic's communications. The worry is that complaints have shifted from "which Claude model is best" to "should I switch to GPT or go local", a sign of eroding brand loyalty rather than healthy competition among models.

RTK is a Rust CLI proxy that cuts 60-90% of terminal token noise before it reaches your coding agent's context window: Coding agents like Claude Code waste large chunks of their context window on repetitive build logs, progress bars, and boilerplate output. RTK (Rust Token Killer) wraps common commands like git status or cargo test, filters the noise, and returns only useful output, keeping the agent's reasoning focused and sessions longer. It is a single Rust binary with no dependencies and is open source under Apache 2.0.

Why Go Is an Ideal Language for AI-Assisted Software Engineering: A Google Developers Blog post argues that as AI agents write more code, the bottleneck shifts from writing to reviewing, and Go's explicit style, mandatory formatting, strong standard library, built-in tooling (fuzzing, profiling, tracing), and strict backward compatibility make AI-generated Go easier to audit and maintain than most alternatives. Commenters push back hard: some cite Tencent's AutoCodeBench multilingual benchmark, which ranks Elixir near the top for LLM-generated code quality, while others argue that languages with stronger type systems (Rust, Haskell) or larger training corpora (JS, Python) are better fits for agentic workflows.

Nvidia's Risky Business: Stratechery's Ben Thompson draws a historical parallel to Jay Cooke's 1870 Northern Pacific Railway gamble to map out Nvidia's compounding risks: CUDA's software moat is real but brittle, hyperscalers are investing heavily in custom silicon, and China exposure adds geopolitical fragility. The Hacker News discussion digs into whether Google's TPU strategy could finally crack CUDA's grip, and what it would take for an alternative to achieve the same ecosystem depth.