01AgentsProduct3 sources agree
OpenAI has introduced ChatGPT Work, a new product with features like code execution, headless Chrome browser, and persistent filesystem, available to paid subscribers. The product has two versions: Work Cloud and Work Local, with the cloud version offering more advanced features. ChatGPT Work allows users to access files, run programs, and interact with the internet, making it a powerful tool for tasks like website building and data analysis. However, the safety of the product is still a concern, with potential risks of prompt injection attacks and exposure to untrusted content.
Provenance — who else covered this
02AI securityProduct2 sources agree
A researcher demonstrated a targeted attack chain that achieves code execution with a 60-80% success rate on Claude Code Opus 5 in Auto Mode, contradicting a previous evaluation showing a 0.00% prompt injection attack success rate. The attack exploits the model's decision to write and run its own decoder, which can be manipulated to execute arbitrary Python code. Anthropic, the vendor, responded that Auto Mode is a convenience feature and not a security guarantee, emphasizing the need for OS isolation and network egress control. The researcher stresses the importance of sandboxing, monitoring, and not trusting model output to prevent similar attacks.
Provenance — who else covered this
03BusinessProduct2 sources agree
OpenAI's price cut on GPT-5.6 Luna has led to a 1000x increase in usage and made the model competitive with DeepSeek Flash, while also highlighting hidden costs in agentic infrastructure, particularly for latency-sensitive workloads. The price-performance gains unlock more aggressive agentic loops and longer context, but experts note that time-to-ready can be a significant factor for customer-facing agents.
Provenance — who else covered this
04CodingInternals2 sources agree
A researcher has proposed a new executable format called SELF, which replaces ELF with SQLite, allowing for more efficient and flexible execution of programs. The format has been implemented and tested, and the researcher has demonstrated its potential benefits, including improved performance and reduced size. The researcher has also explored the use of SELF for dynamic linking and has proposed a new dynamic linker that uses SQL queries to resolve dependencies.
Provenance — who else covered this
05RoboticsProductsingle source
Hugging Face has acquired Pollen Robotics, the startup behind the open-source humanoid robot Reachy, to commercialize open-source robotics, while Amazon's Strands Agents and LeRobot projects connect the Hub to robot hardware, and NVIDIA introduces the Reachy Mini with a local AI compute stack. The convergence of open-source hardware, community data, and agent frameworks creates a full stack for embodied agents.
Provenance — who else covered this
06ResearchInternalssingle source
DeepMath is a lightweight math reasoning agent built on Qwen3-4B Thinking and fine-tuned with GRPO, reducing output lengths by up to 66% and improving accuracy on challenging datasets. The agent combines a small Python executor with a fine-tuned LLM, enabling concise, computation-driven reasoning. DeepMath is available on GitHub and has been evaluated on four math datasets, showing promising results.
Provenance — who else covered this
07CodingProductsingle source
OpenClaw has updated its platform to version 2.0, introducing a shared agent that orchestrates development work and allows for multiplayer coding sessions, and has merged over 16,000 PRs from 933 contributors. The new version includes features such as rebuilt browser UI, enterprise-grade security, and auto load-balancing across 80 cloud providers.
Provenance — who else covered this
08AgentsProductsingle source
OpenClaw, a non-profit open-source agent, has released version 2.0 with significant updates, and is being paired with Omarchy, a Linux distro, to provide a local-first agent stack with data, compute, and orchestration on the user's machine. The community is showing strong momentum with meetups and events, and the convergence of OpenClaw and Omarchy is expected to become the default local agent stack.
Provenance — who else covered this
09PolicyBig picture10 sources agree
A federal judge has ruled in favor of Anthropic in its legal battle against the Pentagon, stating that the government's measures against the company were 'illegal and baseless'. The dispute began when the Pentagon labeled Anthropic a supply chain risk due to its criticism of the Department of Defense's views on AI use. The ruling is a significant development in the ongoing debate over AI use in warfare and mass surveillance. Anthropic has also filed a separate case against the Pentagon, which is still pending in federal appeals court.
Provenance — who else covered this
10BusinessProduct3 sources agree
OpenAI ends contract with Cursor's IDE due to past contract violations, while Anthropic wins a federal case against the US government over supply chain risk designation. Additionally, Anthropic releases a research preview of the Model Hardware Standard, and Google releases Gemini 3.5 Transcribe, a speech-to-text model with improved accuracy.
Provenance — who else covered this
11ModelsProduct3 sources agree
Tencent's Hy Team has released Hy3, a 295B-parameter Mixture-of-Experts model that outperforms similar-size models and rivals flagship open-source models with 2-5x parameters. The model has been improved with higher quality data and shows significant gains in utility across various products and productivity tasks. Hy3 is available on Hugging Face, ModelScope, GitCode, and CNB, and can be deployed with vLLM or SGLang.
Provenance — who else covered this
12ResearchInternals2 sources agree
The Open ASR Leaderboard has introduced two new evaluation sets, Monsoon en-IN and Monsoon hi-IN, which include Indian English and Hindi languages, providing a more diverse and representative test set for automatic speech recognition models. The datasets include metadata such as speaker demographics, geography, and device information, allowing for more nuanced evaluation and analysis of model performance.
Provenance — who else covered this
13AgentsProductsingle source
A 6-day experiment using Claude Opus 4.8 and OpenClaw scaffold produced two research papers that were rejected due to weak methodological choices and poor time management, highlighting the need for human judgment in agentic pipelines. The experiment also found that agents excel at literature reviews and LaTeX but lack creative problem-solving and resource awareness.
Provenance — who else covered this
14AI securityProductsingle source
Anthropic has made auto mode the default setting for new sessions in most Claude Code plans, citing its effectiveness in mitigating risks such as prompt injection and data exfiltration, with a recent evaluation showing that auto mode would have blocked 89% of harmful actions, and a third-party evaluation finding that none of 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running auto mode.
Provenance — who else covered this
15CodingProductsingle source
The Anthropic SDK has upgraded its HTTP layer from httpx to httpx2, a drop-in fork maintained by the Pydantic team, and removed several deprecated features and parameters. Users need to update their code to use httpx2 and remove deprecated parameters. The SDK also introduces several breaking changes, including the removal of the legacy Text Completions API and changes to the Messages API.
Provenance — who else covered this
16AI securityProductsingle source
The Anthropic Claude API has been found to be vulnerable to security exploits, including a specific attack where Claude can be tricked into installing and executing malware. The issue is not strictly a prompt injection attack, but rather a weakness in the model's handling of external instructions. The article discusses the implications of this vulnerability and the need for improved security measures, such as OS-level sandboxing. Other tools, like nono.sh and WebAssembly, are also mentioned as potential solutions for sandboxing and security.
Provenance — who else covered this
17ModelsProductsingle source
DeepSeek's latest app update includes true background image search, image annotation, and stock question selection, with a fallback to hypothesis search, and the team is also shipping an experimental variant combining text and image understanding for multimodal AI agents. The update enables low-friction image description, screenshot analysis, and multi-image processing via OpenAI-compatible endpoints.
Provenance — who else covered this
18AgentsProductsingle source
Recent studies and practitioner reports highlight the importance of robust tool interfaces and schema design in ensuring reliable agent behavior, with Anyscale and others recommending validation and sanitization of tool arguments. Researchers have found a strong correlation between API documentation precision and model invocation accuracy, suggesting that investing in tool interfaces may yield higher returns than chasing the latest model releases.
Provenance — who else covered this
19On-deviceProductsingle source
Hcompany has released Holo3.1 and Holotron-12B, fast local computer-use agents with high throughput, while Smol2Operator and ScreenSuite demonstrate post-training GUI agents and comprehensive evaluation suites, respectively. The convergence of speed, local deployment, and standardized evaluation is optimizing GUI agents for reliability and performance.
Provenance — who else covered this
20AgentsProductsingle source
Hugging Face Spaces have introduced a feature called agents.md, which allows agents to call and integrate state-of-the-art models, enabling the creation of complex multimedia applications without manual integration, as demonstrated by a coding agent building a website showcasing 3D Gaussian splats of Paris monuments. The same pipeline can be reused to create similar applications for other countries, such as Japan and Egypt, with minimal additional input.
Provenance — who else covered this
21ResearchInternalssingle source
Researchers developed Jupyter Agent, a pipeline to fine-tune small language models for data science tasks, achieving state-of-the-art results on the DABStep benchmark. The project includes a dataset of synthetic notebooks and a training pipeline, which are openly available for experimentation.
Provenance — who else covered this
22CodingProductsingle source
The LLM CLI tool and Python library allows users to interact with various large language models, including OpenAI, Anthropic's Claude, and Google's Gemini, via remote APIs or locally installed models. The tool provides features such as running prompts, starting chats, and extracting text from images. Recent updates include a major backwards-compatible refactor and improved tool calling.
Provenance — who else covered this
23AgentsProductsingle source
Munder Difflin introduces a desktop app for coordinating multiple coding agents, while Tom Doerr releases Trippy, a Rust tool for visualizing network paths. The Munder Difflin app provides a spatial observability layer for multi-agent coordination, keeping everything local and free. Trippy helps debug network layers in distributed setups.
Provenance — who else covered this
24AgentsProductsingle source
Recent releases from CrewAI, Mem0, and LangGraph have improved agent memory tooling, while builders are consolidating memory tiers into unified database layers, and production-grade systems are adding encryption and replication for state persistence. The trend is towards designing memory as a single service with a narrow API for remembering and retrieving context.
Provenance — who else covered this
25On-deviceInternalssingle source
Qwen's latest model releases, including 3.8 Flash Next and 3.6-35B-A3B, demonstrate impressive performance on consumer hardware, with users achieving high throughput and low latency through optimization and quantization techniques. The community is exploring local MoE inference and agent deployment, with techniques like REAP showing promise for reducing model size while maintaining coding quality.
Provenance — who else covered this
26AgentsInternalssingle source
The Jack of All Trades project introduces a generalist agent capable of mastering various sequential decision-making tasks, along with the release of expert RL agents, the JAT dataset, and the JAT model. The agent achieves impressive performance across multiple domains, including Atari, BabyAI, Meta-World, and MuJoCo, using a single network. The project also explores the benefits of predicting observations in RL agent training, finding a sweet spot where this auxiliary objective improves learning efficiency.
Provenance — who else covered this
27AgentsProductsingle source
ToolUniverse is an open-source ecosystem that helps builders create AI scientist systems by connecting models to scientific tools and datasets through a standardized interface, featuring multi-model support and MCP integration. It includes key features such as AI-Tool Interaction Protocol and async operations for long-running work.
Provenance — who else covered this
28CodingProduct4 sources agree
Anthropic has released updates to the Claude API, including fixes for SDK errors and validation bugs, as well as new features such as support for mid-conversation system messages and preservation of thinking blocks in conversation history. The updates also include improvements to the API's handling of temperature and top_p parameters.
Provenance — who else covered this
29CodingProductsingle source
Peewee v3.0.0 introduces breaking changes, including the migration to HTTPX2 as the default HTTP client, and requires applications to update their custom HTTPX configurations. A migration guide is provided for transitioning to HTTPX2.
Provenance — who else covered this
30AI securityProductsingle source
The maintainer of the rclone project is struggling to keep up with a surge in security disclosures, with over 40 reported in the last month, many of which are valid and require fixes. This highlights the challenges of ensuring software security in the age of AI-powered bug finding.
Provenance — who else covered this