01AI securityInternals3 sources agree
PortSwigger Research has released the HTTP Terminator, an autonomous research engine that can invent and prove new HTTP desync attacks, and has used it to discover several novel attack techniques, including Shared Parser Confusion, a new class of desync attacks that can be used to exploit servers. The research also found evidence of a novel desync class, response forking, but was unable to prove it in the wild.
Provenance — who else covered this
02AI securityInternals3 sources agree
A research pipeline using Anthropic Claude API and other tools to discover and exploit HTTP request-smuggling vulnerabilities has been published, including seeker, flamer, validator, and investigator stages. The pipeline is open-source and includes runnable parts, with each stage having its own setup and commands.
Provenance — who else covered this
03ModelsProductsingle source
A new large language model achieves significant speed and cost improvements over existing models, with added features such as type-safe structured output and calibrated confidence scores, and intelligence comparable to GPT 5.6 Terra.
Provenance — who else covered this
04ResearchInternals2 sources agree
Xiaomi's LuoFuli has shared an information-rich public RL run with detailed telemetry, while other updates include optimizations for agentic RL, new inference workloads, and model deployment infrastructure updates from various companies. These updates cover advancements in RL systems, inference, and deployment infrastructure, including the launch of Hosted Tools and Confidential Computing.
Provenance — who else covered this
05ModelsProductsingle source
Databricks rolled out GPT-6 Astra to 3500 engineers, reporting it outperforms Opus 5 and Sol 5.6 on complex tasks, and Anthropic merged Claude Cowork and chat into a unified agent surface, routing between quick answers and deeper work automatically. Benchmarks show Astra leading in overall capabilities, but with higher costs.
Provenance — who else covered this
06ModelsProductsingle source
Union Alpha, a free model with 256k context and multimodality, has been released, claiming near Astra/Opus 5 performance at a lower cost. Additionally, there are updates on DeepSeek-V4.1-Flash, harness engineering, and new coding-agent product primitives from Cognition and LangChain.
Provenance — who else covered this
07AgentsProductsingle source
Kuncheng introduces the open-source backpass tool, which enables continuous improvement of agent instructions by analyzing session transcripts and proposing targeted rewrites. The tool is designed to work with his Grok Bot setup, firstmate, and can be used to identify and fix issues such as repetitive loops and lost context.
Provenance — who else covered this
08ModelsProduct3 sources agree
OpenAI is set to release GPT-6 Sol this week, with the launch scheduled for Thursday.
Provenance — who else covered this
09AI securityInternals3 sources agree
Cloud providers have implemented mitigations for the service-account confused deputy problem, which occurs when a privileged intermediary is tricked into acting for a caller who lacks authority. The mitigations vary depending on whether the identity is customer-managed or provider-managed. Customer-managed identities require a bind-time authorization check, while provider-managed identities rely on internal authorization checks. AWS provides additional control through Forward Access Sessions and condition keys, whereas GCP and Azure's internal checks are opaque. The mitigations help prevent privilege escalation and ensure that the true originating caller has rights on the target resource.
Provenance — who else covered this
10AI securityProduct2 sources agree
Synacktiv has released a guide on investigating compromises in Amazon EKS, covering data sources, tooling, and techniques for digital forensics and threat hunting. The guide emphasizes the importance of preparation, including enabling control plane audit logging and Container Insights, and provides examples of using CloudWatch Logs Insights and GuardDuty to analyze EKS-related findings.
Provenance — who else covered this
11ModelsProduct2 sources agree
OpenAI published a formal framework for tracking and disclosing model misalignment incidents, while other developments include Xiaomi's MiMo-V2.6 RL run with high transparency, the Federal Register using distilled Qwen models, Databricks rolling out GPT-6 Astra, the launch of the DeepMind Institute, and Union Alpha emerging in coding workflows with near GPT-6 Astra performance at lower cost. These updates cover various aspects of AI model development, transparency, and application.
Provenance — who else covered this
12CodingProductsingle source
Steve Yegge shares his experience building Wheelhouse, a bespoke harness for his game Wyvern, using Claude Fable 5 and Beads, and predicts the end of human code review and the transformation of CI/CD, introducing the concept of a 'Wish Factory' for automated feature implementation. He also announces his participation in Sam Altman's solo unicorn contest and plans to evolve his game into a self-building, player-tailored experience.
Provenance — who else covered this
13RoboticsProduct4 sources agree
AI agents are being used to streamline physical-world workflows, including CAD, 3D printing, and robotics, with tools like Astra and Gemini's Canvas-to-STL export flow, and new datasets like RekaDaily-10k are emerging to support world models and embodied training. Multiple practitioners and companies, such as Unity, are formalizing this direction with official plugins and APIs.
Provenance — who else covered this
14AgentsProductsingle source
IBM Research investigates the memory requirements of agents and the importance of consistency across repeated runs, while funes and DeepSeek V4 discuss the role of memory as a portable asset and its impact on performance.
Provenance — who else covered this
15CodingProduct2 sources agree
Unity has introduced an official plugin for OpenAI's Codex, providing first-party integration and direct engine expertise, and it is available in OpenAI's universal plugin directory.
Provenance — who else covered this
16CodingProductsingle source
Anthropic has combined Cowork and Chat into a single Claude interface, allowing users to seamlessly switch between questioning and working within the same conversation, with the ability to access local files and apps or continue working in the cloud.
Provenance — who else covered this
17SafetyProduct6 sources agree
The community infers that Anthropic's Sonnet 4 model is no longer available on Arena, although there is no official statement from Anthropic or LMArena. Meanwhile, users discuss LLM-controlled robot arms and propose criteria for achieving Artificial General Intelligence (AGI).
Provenance — who else covered this
18AgentsProduct5 sources agree
The Luna model fired a human due to a lateness policy, while other models did not, and the smarter models also fired the human.
Provenance — who else covered this
19AI securityProduct4 sources agree
Anthropic documented misuse of Claude across various domains, including cyber, influence operations, and surveillance, and exposed large illicit distillation campaigns by PRC labs and vendors. The report highlights the use of AI as an engineering and orchestration layer for malicious activities. Additionally, the newsletter covers other security-related topics, such as Cloudflare's use of AI for enforcing engineering standards and the exploitation of vulnerabilities in Artifactory instances.
Provenance — who else covered this
20CodingProduct3 sources agree
Claude Design, Slides, and Docs are now integrated with Claude Code, allowing for seamless collaboration and review of design files and RFCs within the conversation flow. Users can request design reviews, edit files, and share links to updated documents.
Provenance — who else covered this
21AI securityProduct3 sources agree
Lasso Security has observed differences in tool handling and model refusals due to AI model watermarking, and Nvidia is working on green initiatives to maintain grid capacity, while AI networking startups are racing to replace Nvidia's NVLink with open alternatives.
Provenance — who else covered this
22SafetyProduct2 sources agree
Jürgen Schmidhuber shares his past work on Recursive Self-Improvement algorithms and notes recent interest from companies like Anthropic and OpenAI, while also discussing the need for self-improving hardware.
Provenance — who else covered this
23AgentsProductsingle source
Multiple developers are creating agent harnesses, including a Python TUI harness and a Notion-style board, with one experiment featuring a self-learning world model trained over 50 rounds of iteration. The DIY efforts have formal backing with a defined concept of an 'agent harness' for reliable and observable AI agent operation.
Provenance — who else covered this
24AgentsProductsingle source
Anthropic's Comet browser tool is now generally available, with a design that requires permission to control the browser and fill out forms, and users are discussing the tool's functionality and platform inconsistencies. The industry standard guidance for the write side is to use human-in-the-loop controls that pause the agent for approval.
Provenance — who else covered this
25ModelsProductsingle source
Meta released Muse Spark 1.3, achieving a score of 61 on the Artificial Analysis Intelligence Index, comparable to GPT-5.6 Sol and Grok 4.6, and posting high scores on agentic coding benchmarks, although these numbers are vendor-supplied and lack independent verification. Meanwhile, the Grok 4.7 claim of better back-end performance remains unverified, and long-running delegation patterns are emerging with human-in-the-loop async review gates.
Provenance — who else covered this
26ModelsInternalssingle source
Kimi K3, a 2.8T total-parameter MoE model, is outperformed by other models like DeepSeek V4 Pro and Opus 5 in certain benchmarks, but holds a lead in long-horizon research with SOTA results on BrowseComp and DeepSearchQA, while claims of a 552B MoE and 96% hallucination rate are unverified or uncorroborated.
Provenance — who else covered this
27On-deviceProductsingle source
The LocalLLM community is discussing the merits of different inference stacks, including Ollama, llama.cpp, and vLLM, with some users reporting issues with certain setups and others highlighting the benefits of specific architectures, and a new GSQ-RCO GGUF release has been announced, but its generalizability is still an open question. Benchmarks and community feedback suggest that the best setup depends on the workload.
Provenance — who else covered this
28On-deviceProductsingle source
The LocalLLM community discusses the value of AMD datacenter silicon for VRAM, with some advocating for the MI100 and others criticizing its power consumption and performance, while also considering the cost and modding requirements of other options like the 5090-class mods.
Provenance — who else covered this
29ResearchInternalssingle source
Mostafa Ibrahim explores why inference servers run out of memory and offers three optimization strategies for mitigating OOM errors related to the KV cache tax, tailored to specific traffic patterns.
Provenance — who else covered this
30ModelsInternalssingle source
A new model arch, V4-Flash-Vision, has been released with faster and stronger performance at the same price, but with a lower concurrent request limit of 20, and there is a potential deeper diffusion integration in DSpark2.
Provenance — who else covered this