A self-spreading worm compromised 32 Red Hat npm packages in a supply chain attack linked to the same group that breached GitHub, the Agentic AI Foundation added its first new hosted project, and NVIDIA dropped an open foundation model for physical AI. Here’s what matters.
Red Hat npm packages compromised by “Miasma” — a self-spreading credential-stealing worm
Wiz Research disclosed on June 1 that 32 packages under Red Hat’s @redhat-cloud-services npm namespace were compromised across 96 versions, collectively downloaded ~117,000 times per week. The malware, dubbed Miasma, is a self-propagating worm that uses stolen npm tokens to republish backdoored versions of other packages — it sweeps for GitHub Actions tokens, AWS/GCP/Azure credentials, Vault tokens, SSH keys, Docker registry credentials, and every .env file it can find. The attack used a compromised Red Hat employee GitHub account to push malicious orphan commits to RedHatInsights repositories, bypassing code review. The tradecraft traces back to TeamPCP, the same group behind the GitHub internal repos breach and earlier compromises of Trivy, Checkmarx, and Bitwarden CLI — all through developer tooling. Red Hat says the packages were limited to internal development and no customer environments were affected, but the pattern is alarming: TeamPCP is systematically targeting the developer supply chain, and each compromised tool gives them a foothold to reach the next one.
Agentic AI Foundation adopts agentgateway as its newest hosted project
The Agentic AI Foundation (AAIF), the Linux Foundation entity that hosts MCP, goose, and AGENTS.md, welcomed agentgateway on June 4 — an open-source gateway purpose-built for MCP, Agent-to-Agent (A2A), LLM, and API traffic. The project gives organizations a single chokepoint for policy enforcement, rate limiting, and observability across all their agentic AI traffic, donated to a vendor-neutral home under the Linux Foundation. This is AAIF’s first project adoption since its December 2025 founding, and it signals that the foundation is moving beyond its initial contributions toward building out the infrastructure layer that enterprises need before they can deploy agents at scale. With MCP Dev Summits in Bengaluru (June 9-10) and Mumbai (June 14-15) coming up alongside KubeCon India, AAIF is building momentum quickly.
NVIDIA launches Cosmos 3 — an open foundation model for physical AI
Announced at Computex on June 1, Cosmos 3 is NVIDIA’s open world foundation model that unifies physical reasoning, world generation, and action prediction in a single system. The release includes two variants: Cosmos 3 Nano (16B parameters, targeting workstation-class RTX PRO 6000 hardware for real-time robotics inference) and Cosmos 3 Super (64B parameters for datacenter Hopper/Blackwell deployments). Trained on 20 trillion tokens of multimodal data — nearly a billion images, 400 million real and synthetic videos, audio, text, and robotic action data — NVIDIA is shipping training scripts, datasets, and deployment tools alongside the model. They also launched the Cosmos Coalition with Black Forest Labs, Runway, Skild AI, and others to advance open world models. Coming the same week as the Physical AI Agent Toolkit at GTC Taipei, NVIDIA is making a clear play to own the open-source stack for robotics and autonomous systems end-to-end.
AMD GAIA 0.20 ships with per-agent device selection and multi-device AI
AMD’s open-source framework for building AI agents that run entirely on local hardware hit version 0.20 with a significant feature upgrade. The headline addition is per-agent device selection — each AI agent can now declare which accelerators it supports (CPU, GPU, or NPU) via DeviceConfig tuples, and users can select a device per agent through the UI or a new --device CLI flag. Other additions include tighter security controls over what agents can access on the system, an Agent Hub TUI for terminal-native agent management, PowerPoint RAG support, and up to 10x faster email organization. GAIA is interesting as a counterweight to the cloud-centric agent story: it’s betting that privacy-sensitive and air-gapped use cases will drive demand for fully local agent execution, and AMD is using it to showcase that its Ryzen AI / Instinct hardware can run the stack without NVIDIA.
Blender 5.2 LTS enters beta — first long-term support release of the 5.x series
Blender 5.2 hit beta this week, and it’s the first LTS release in the 5.x series — meaning it’ll receive two years of critical fixes after its July 8 target GA date. Notable features include EEVEE instancing optimizations (up to 2x faster in CPU-bottlenecked scenes), support for online asset libraries where remotely-hosted libraries can be browsed and downloaded on demand, and new input color spaces for Apple, ARRI, Blackmagic, Canon, and Sony cameras. The LTS designation matters for studios and production pipelines that need stability guarantees, and Blender reaching 5.2 LTS while maintaining its fully open-source GPL model continues to be one of the strongest arguments that community-governed open source can produce industry-leading creative tools.
The annual State of Open Source Report from Perforce OpenLogic, OSI, and the Eclipse Foundation landed this week with data from 700+ respondents. The topline: less than 2% of organizations decreased their open-source consumption. But the operational reality is grimmer — 60% of respondents at large enterprises (5,000+ employees) spend at least half their time on maintenance, production issues, and bug fixes rather than feature development. Avoiding vendor lock-in surged 22 percentage points year-over-year, now cited by 55% of respondents, reflecting the relicensing wave that rattled the industry. The report frames open source as no longer just an engineering preference but a “strategic asset and strategic responsibility” shaped by geopolitical pressure, security risk, and compliance complexity — a framing that aligns with everything from the EU CRA to IBM’s Project Lightwell investment.