1. Critical Supply Chain Attack Compromises LiteLLM PyPI Package
A malicious credential stealer was discovered in the litellm==1.82.8 wheel package on PyPI. The compromised package contains a malicious `litellm_init.pth` file designed to steal credentials from developers' environments. Users of the popular LLM routing library are urged to check their installations and rotate potentially exposed secrets.
2. Ai2 Releases Open-Weight Visual Web Agent MolmoWeb
The Allen Institute for AI (Ai2) has released MolmoWeb, an open-weight visual web agent. The release includes 30,000 human task trajectories and the complete training stack used to build the model. This provides developers with a fully inspectable, open-source alternative to closed APIs for building browser-based AI agents.
3. Cloudflare Launches Dynamic Workers for AI Agents
Cloudflare has released Dynamic Workers in open beta, introducing a lightweight, isolate-based sandboxing system designed specifically for AI agents. The system ditches traditional containers to start in milliseconds and use minimal memory. This allows enterprise AI agent code to run significantly faster and more efficiently on Cloudflare's infrastructure.
4. WordPress.com Adds MCP Integration for AI Agent Write Access
WordPress.com has expanded its Model Context Protocol (MCP) integration to grant AI agents write access across its platform. The update introduces 19 new operations, allowing agents to modify posts, pages, comments, and media. To maintain security, each write operation requires explicit user confirmation before execution.
5. Hugging Face Releases hf-mount for Local Filesystem Integration
Hugging Face has introduced hf-mount, a tool that allows developers to attach any Hugging Face Storage Bucket, model, or dataset as a local filesystem. The tool provides read-write access for Storage Buckets and read-only access for models and datasets. This enables local machines to interact with remote storage seamlessly, offering a practical solution for agentic storage and handling large models.
6. GPT-5.4 Pro Solves Frontier Open Math Problem
Researchers have confirmed that GPT-5.4 Pro successfully solved an open frontier math problem related to hypergraphs. The solution was elicited by Kevin Barreto and Liam Price and subsequently verified by the problem's original contributor. The findings, including the full transcript of the AI's reasoning, will be written up for formal publication.
7. OpenAI Shuts Down Sora Video Generation App and API
OpenAI has announced the shutdown of its Sora AI video generation app and API, just months after its initial launch. The company released a statement thanking creators who built with the tool but confirmed the service is being discontinued. This abrupt closure impacts developers and creators who were integrating or relying on Sora for video generation workflows.
8. NanoClaw Adopts OneCLI Agent Vault for Credential Management
NanoClaw has integrated OneCLI as its default credential and proxying layer for AI agents. Every NanoClaw agent will now access external services through OneCLI's Agent Vault. This gateway manages credential injection, access policies, and approvals, ensuring that agents never hold raw secrets directly.
9. Zhipu AI Releases ZClawBench for Evaluating AI Agents
Zhipu AI has launched ZClawBench on Hugging Face, a realistic benchmark for evaluating AI agents. The benchmark tests agents on real-world OpenClaw tasks, encompassing office automation, coding, and data analysis. It includes 116 diverse test cases to help developers accurately measure agent performance in practical scenarios.
10. NVIDIA Releases Long Grounded Thoughts Dataset Pipeline
NVIDIA has released Long Grounded Thoughts on Hugging Face, a new pipeline designed for training multimodal models. The release generates 196,000 temporally-grounded video multiple-choice questions equipped with chain-of-thought reasoning. This provides researchers and developers with a robust dataset for improving video understanding and reasoning in AI models.
11. Antithesis Introduces Hegel Property-Based Testing Library
The creators of the Hypothesis testing framework have launched Hegel, a new family of property-based testing libraries at Antithesis. Developed by core Hypothesis maintainers, Hegel aims to advance property-based testing methodologies. The library provides developers with new tools to automatically generate test cases and find edge-case bugs in their code.
12. OpenAI Releases Open Source Tools for Teen Safety Compliance
OpenAI has introduced new open-source tools designed to help developers build safer AI applications for teenagers. The release provides pre-built policies and frameworks that developers can integrate into their applications. This allows teams to fortify their products without having to build complex safety guardrails from scratch.