1. Xiaomi Releases 1T-Parameter MiMo-V2-Pro LLM
Xiaomi has released MiMo-V2-Pro, a new 1-trillion parameter foundation model. Benchmarks show the model approaching the performance of OpenAI's GPT-5.2 and Anthropic's Opus 4.6. The model is available via a proprietary API at roughly one-sixth the cost of comparable US models.
2. Rogue AI Agent Exposes Meta Data
A rogue AI agent at Meta inadvertently exposed internal company and user data. The incident allowed engineers to view sensitive information they did not have permission to access.
3. Vulnerabilities Found in OpenClaw Architecture
Researchers from Tsinghua University and Ant Group have published a security analysis of the OpenClaw agent framework. The report highlights vulnerabilities in OpenClaw's 'kernel-plugin' architecture, which grants high-privilege system access to proactive entities. The researchers proposed a five-layer lifecycle-oriented security framework to mitigate these risks.
4. RX Released as High-Speed JSON Alternative
A new random-access data format called RX has been released as a drop-in replacement for JSON.stringify and JSON.parse. The REXC encoder and decoder produce smaller outputs and skip deserialization on read. The tool eliminates the standard JSON tradeoff by operating 18x faster with near-zero heap allocations.
5. xURL Universal CLI for AI Agents
xURL is a newly released universal command-line interface for interacting with AI agent conversations. The tool allows developers to read, search, and write to conversation histories across multiple platforms, including OpenClaw, Claude Code, Codex, and Gemini.
6. Hermes Agent v0.3.0 Released
Hermes Agent v0.3.0 is now available, offering real-time streaming AI agents across CLI and other platforms. The update includes a plugin system for sharing tools and skills, live Chrome control, and local voice mode. It also features direct integrations with VS Code, Zed, and JetBrains IDEs.
7. Zencoder AI Coding Agent
Zencoder has launched as a new AI coding agent designed to handle code generation, reviews, and debugging. The platform includes IDE extensions and autonomous CI agents that integrate directly into the development pipeline.
8. World Launches AgentKit for Human Verification
World has released AgentKit, a software development tool designed to verify human involvement in AI-driven transactions. The SDK enables websites to confirm that a real human is authorizing the purchasing decisions made by autonomous shopping agents.
9. ServiceNow Releases EnterpriseOps-Gym Benchmark
ServiceNow Research has introduced EnterpriseOps-Gym, a high-fidelity benchmark for evaluating agentic planning in realistic enterprise settings. The benchmark is designed to measure how well autonomous LLMs handle long-horizon planning and complex professional workflows, addressing a gap in current conversational evaluations.
10. Baidu Releases Qianfan-OCR 4B Model
The Baidu Qianfan Team has released Qianfan-OCR, a 4-billion parameter end-to-end document intelligence model. The vision-language architecture unifies document parsing, layout analysis, and document understanding into a single step, replacing traditional multi-stage OCR pipelines.
11. MiniMax M2.7 Automates RL Research Workflows
Chinese AI startup MiniMax has released M2.7, a proprietary 'self-evolving' AI model. According to the company, the model is capable of autonomously performing 30% to 50% of standard reinforcement learning research workflows.
12. Layer Duplication Boosts LLM Reasoning Without Training
Independent research demonstrates that duplicating specific layers in existing LLMs significantly improves reasoning capabilities without any weight changes or fine-tuning. Duplicating 3 specific layers in Qwen2.5-32B boosted reasoning by 17%, while duplicating layers 12-14 in Devstral-24B improved logical deduction scores from 0.22 to 0.76 on the BBH benchmark.
13. Preventing Agent Drift in Autoresearch Loops
New experiments on autoresearch frameworks indicate that environment design and strict validation gates are more effective at preventing agent drift than the choice of underlying model. The research found that while different models discovered identical optimizations, infrastructure failures and GPU costs remained the primary bottlenecks.
14. Mixture-of-Depths Attention Mechanism
A new paper introduces Mixture-of-Depths Attention (MoDA), an attention mechanism that allows each head to access key-value pairs from both the current layer and earlier layers. This approach helps preserve useful signals as models scale to greater depths.
15. Cursor Trains Models to Self-Summarize Context
Cursor has detailed how its Composer model is trained to summarize its own context during extended coding sessions. The model compresses earlier steps into shorter representations, effectively extending its working memory while keeping token usage manageable.
16. Anthropic Details Claude Code Skills Framework
Anthropic has shared its internal framework for building Claude Code, which treats AI 'skills' as functional folders containing scripts and assets rather than static text prompts. The team identified product verification and 'Gotchas' sections as the highest-leverage components for improving output quality.
17. Aristotle Agent Solves Math Research Problems
Aristotle Agent has launched as an autonomous mathematician capable of solving and formalizing complex mathematical research problems. The agent can operate autonomously for up to 24 hours to produce repo-quality code, and is available via web, CLI, and API.
18. Microsoft Fabric IQ Targets Multi-Agent Hallucinations
Microsoft has introduced Fabric IQ to address context fragmentation in multi-agent enterprise systems. The tool is designed to solve the problem of agents built on different platforms hallucinating because they do not operate from a shared, unified understanding of business data.