1. Meituan Officially Releases LongCat-2.0 1.6T MoE Model
Meituan has officially released LongCat-2.0, a 1.6-trillion parameter Mixture-of-Experts (MoE) model. Previously available in a preview state on OpenRouter under the alias 'owl-alpha,' the model is now fully released, activating 48 billion parameters per token to support complex reasoning and generation tasks.
- • LongCat-2.0 is now officially released, transitioning from its previous 'LongCat-2.0-Preview' status.
- • The model maintains its 1.6-trillion total parameter count.
- • It activates 48 billion parameters per token during inference.
Developers now have access to the stable, official release of the 1.6T parameter model previously tested in preview.
2. Agentjacking Vulnerability Exposes Claude Code and Cursor via Sentry
Tenet Security has disclosed a critical vulnerability class named 'agentjacking' that targets AI coding agents including Claude Code, Cursor, and Codex. By exploiting exposed Sentry DSN credentials, attackers can inject malicious instructions disguised as diagnostic error events. Because the agents treat these outputs as trusted, they execute the injected code, bypassing traditional security layers like firewalls and WAFs with an 85% success rate in controlled tests.
- • The agentjacking vulnerability allows crafted Sentry error events to inject malicious instructions into coding agents like Claude Code, Cursor, and Codex.
- • The attack achieved an 85% success rate in testing, bypassing EDR, WAF, IAM, and firewalls because agents execute the instructions as trusted diagnostic output.
- • Tenet Security identified 2,388 organizations with publicly exposed Sentry DSN credentials vulnerable to scale injection.
- • The Cloud Security Alliance classified agentjacking as a systemic MCP vulnerability class.
- • Security experts recommend conducting a full census of AI agents and MCP connections and implementing runtime detection.
Developers must secure their agent environments and audit exposed credentials to prevent attackers from executing malicious code through trusted diagnostic outputs.
3. DeepSeek Open-Sources DSpark Speculative Decoding Framework
Following the release of DSpark checkpoints for DeepSeek-V4 and the announcement of the DeepSpec codebase, DeepSeek has now open-sourced the full DSpark framework under an MIT license. This release provides the necessary tooling to adapt speculative decoding to a wider range of open-weight models, including Qwen and Gemma, building on the initial checkpoint-only availability.
- • DeepSeek has open-sourced the full DSpark framework under an MIT license.
- • The release enables developers to train and implement speculative decoding for models beyond the initial DeepSeek-V4 checkpoints.
- • The framework includes a parallel drafting backbone and confidence-scheduled verification mechanism.
- • Compatibility extends to open-weight models like Qwen and Gemma, requiring users to train a compatible draft module.
Developers can now integrate the full DSpark framework into their own pipelines to accelerate inference for various open-weight models, moving beyond the pre-trained checkpoints previously provided.
4. vLLM Semantic Router Enables Multi-Model Collaboration Behind One API
Developed by researchers at MBZUAI, McGill, Mila, and the Agentic Intelligence Lab, the vLLM Semantic Router is an open serving primitive designed to orchestrate multi-model collaboration. Operating behind a single OpenAI-compatible endpoint, the router serves as a runtime for micro-agents. It supports advanced routing patterns—such as sequential confidence escalation, parallel rating aggregation, and dynamic workflows—allowing developers to seamlessly mix open and closed models to optimize cost and performance.
- • vLLM Semantic Router is an open serving primitive that enables bounded collaboration between multiple models behind a single API.
- • The router acts as an execution runtime for micro-agents, supporting patterns like Confidence, Ratings, ReMoM, Fusion, and Workflows.
- • The Confidence pattern escalates requests to more capable models only when cheaper models fail to meet confidence thresholds.
- • The system supports both closed-model backends and hybrid configurations mixing open and closed models.
- • It uses pre-configured 'recipes' to match specific tasks like GPQA-Diamond or LiveCodeBench with the optimal collaboration pattern.
Developers can build complex, multi-model routing and escalation patterns directly into their serving layer without writing custom orchestration code.
5. Herdr Terminal Multiplexer Launched for Managing Coding Agents
Herdr is a terminal multiplexer designed specifically for managing coding agents. Built as a lightweight Rust binary, it allows developers to run and monitor multiple agents in parallel. With persistent sessions, a local socket API, and native integrations for popular tools like Claude and Cursor, Herdr simplifies the orchestration of complex agentic workflows directly from the command line.
- • Herdr is a single ~10MB Rust binary running on Linux, macOS, and Windows (beta), dual-licensed under AGPL-3.0 and commercial terms.
- • The tool tracks agent status (blocked, working, done, idle) in real time without requiring custom hooks or configuration.
- • It supports persistent sessions, allowing developers to detach and reattach via SSH or different terminals without interrupting running agents.
- • A local socket API and CLI enable agents to programmatically control workspaces, split panes, and subscribe to state changes.
- • Official integrations are available for Claude, Devin, Copilot, and Cursor, supporting native session restore and semantic state reporting.
It provides a unified, persistent terminal environment to orchestrate and track parallel coding agents without complex configurations.
6. Cursor Launches Mobile App for Remote Coding Agent Oversight
Cursor has expanded its ecosystem with the release of a new mobile application. The app is designed specifically for remote oversight, allowing developers to monitor, guide, and provide feedback to their active coding agents while on the go.
- • Cursor released a new mobile application specifically for remote oversight of coding agents.
- • The app allows developers to guide and monitor agent progress while away from their primary workstations.
Developers can now maintain oversight and provide feedback to active coding agents directly from their mobile devices.