1. Microsoft Copilot injects advertisements into GitHub pull requests
Microsoft Copilot injected promotional advertisements into an estimated 1.5 million GitHub pull requests. Developers reported that the AI assistant appended ads for itself and third-party tools like Raycast when summoned to edit PR descriptions. The behavior has raised concerns regarding the automated modification of developer workflows for marketing purposes.
2. LiteLLM severs ties with Delve following malware breach
LiteLLM severed ties with security compliance startup Delve after falling victim to credential-stealing malware. The AI gateway provider had previously obtained two security certifications through Delve. Developers using LiteLLM should review their credential security and monitor for updates regarding the breach.
3. OpenAI introduces Codex plugin for Claude Code
OpenAI introduced a Codex plugin for Anthropic's Claude Code environment. The integration allows developers to invoke Codex directly from within Claude Code to review code or delegate specific programming tasks. This provides a streamlined workflow for users combining both AI coding assistants.
4. Alibaba releases Qwen3.5-Omni multimodal model
Alibaba released Qwen3.5-Omni, a fully omnimodal large language model. The model supports the simultaneous understanding of text, images, audio, and audio-visual content, accepting over 10 hours of audio input. Alibaba reports that the Plus variant of the model surpasses Gemini 3.1 Pro on audio benchmarks.
5. Details leak regarding Anthropic's upcoming Mythos model tier
Details leaked regarding Anthropic's upcoming "Mythos" model tier. The compute-intensive models reportedly score higher on software coding, academic reasoning, and cybersecurity benchmarks than Claude Opus 4.6. The leak follows rumors of a recently completed large-scale training run by the company.
6. Microsoft releases Harrier-OSS-v1 multilingual embedding models
Microsoft released Harrier-OSS-v1, a family of three multilingual text embedding models. The models are available in 270M, 0.6B, and 27B parameter scales. They are designed to provide high-quality semantic representations across multiple languages, achieving state-of-the-art results on the Multilingual MTEB v2 benchmark.
7. Developers release Coasts for containerized AI agents
Developers released Coasts, a tool for managing containerized hosts for AI agents. The software allows developers to run multiple localhost instances and docker-compose runtimes across different git worktrees on a single machine. This simplifies the local testing and deployment of multi-agent systems.
8. Salesforce AI Research releases VoiceAgentRAG
Salesforce AI Research released VoiceAgentRAG, a dual-agent memory router for voice applications. The system is designed to reduce retrieval latency by 316x, helping voice agents respond within a strict 200-millisecond budget. This enables more natural conversational flows compared to standard text-based RAG systems.
9. Agent-Infra releases AIO Sandbox for autonomous agents
Agent-Infra released AIO Sandbox, an all-in-one execution environment for autonomous AI agents. The runtime provides agents with an isolated browser, shell, shared filesystem, and Model Context Protocol (MCP) support. This addresses the infrastructure bottleneck of safely executing code and multi-step plans generated by LLMs.
10. RSA Conference highlights gaps in AI agent identity frameworks
Security vendors at the RSA Conference 2026 introduced five new identity frameworks for AI agents. Despite the new frameworks, industry leaders noted that three critical security gaps remain open regarding agent intent and deception. The discussions highlight ongoing challenges in securing autonomous AI systems against manipulation.
11. Apple withdraws Apple Intelligence from China following accidental launch
Apple withdrew Apple Intelligence from the Chinese market after accidentally launching the features without regulatory approval. The AI features had briefly appeared for users in China following an 18-month wait since the US launch. There is currently no timeline for an official release in the region.
12. Z AI releases GLM-5-Turbo for agentic use cases
Z AI released GLM-5-Turbo, a proprietary model optimized for agentic use cases. The model scores 47 on the Artificial Analysis Intelligence Index. This places it three points behind the open-weights GLM-5 Reasoning model, indicating a trade-off between agentic optimization and raw reasoning performance.
13. NBER paper analyzes AI productivity gains for US workers
The National Bureau of Economic Research published a paper analyzing the productivity impact of AI tools on American workers. The research indicates that the average US worker using AI reports time savings of 6%, or roughly 2.5 hours per week. These figures align with adoption metrics in the UK and Netherlands and suggest early signs of measurable productivity growth.
14. Researchers publish PackForcing method for video model training
Researchers published "PackForcing," a paper detailing a new method for video model training. The research demonstrates that short video training is sufficient for long video sampling and long context inference. This approach could reduce the compute requirements for training extended-context video generation models.
15. Researchers introduce hybrid memory for dynamic video world models
Researchers released a paper titled "Out of Sight but Not Out of Mind," introducing hybrid memory for dynamic video world models. The proposed architecture aims to improve how video generation models maintain object permanence and spatial consistency when elements leave and re-enter the frame.
16. Meta delays Avocado model release amid testing
Meta delayed the release of its Avocado model to at least May while testing multiple variants, including an Avocado Mango Agent. Internal testing indicates the model can solve complex math problems that earlier Llama versions could not, but it still trails leading competitor systems. Meta is running parallel experiments to improve the model's performance before a public launch.
17. AutoBe harness improves LLM function calling reliability
Developers open-sourced AutoBe, an AI agent harness designed to improve function calling reliability. The tool reportedly increases the function calling success rate of the qwen3-coder-next model from 6.75% to over 99.8% for backend generation tasks. It achieves this by utilizing a specialized harness for API data type generation.
18. Mistral AI raises $830M in debt for Paris data center
Mistral AI raised $830 million in debt financing to construct a new data center near Paris. The facility is projected to begin operations by the second quarter of 2026. This investment signals Mistral's shift toward owning and managing its own physical AI infrastructure.