Bleeding Llama Flaw & AMD MI350P PCIe Pivot

A critical vulnerability dubbed "Bleeding Llama" (CVE-2026-7482) has been discovered in Ollama, the popular tool for running LLMs locally. The flaw is a heap out-of-bounds read that allows unauthenticated attackers to leak process memory.

Security researchers estimate that over 300,000 servers are currently internet-exposed and vulnerable. The leak can include highly sensitive data, such as system prompts, conversation history, and embedded API keys used by agentic integrations.

AMD has officially launched the MI350P, its first PCIe-based Instinct accelerator in four years. This move targets the "massive middle" of the enterprise market that requires on-premises AI capacity but cannot support liquid-cooled OAM server modules.

The MI350P is designed for air-cooled environments and serves as a drop-in upgrade for standard 2U and 4U servers. This pivot signals AMD's intent to capture the surging demand for private agentic AI deployments within traditional data centers.

At Cloud Next '26, Google announced GKE Agent Sandbox, a new security feature that leverages gVisor to provide a multi-layered defense for AI agents. As agents increasingly generate and execute their own code, the risk of "jailbreak-to-host" attacks has skyrocketed.

The sandbox provides kernel-level isolation, ensuring that even if an agent is compromised via prompt injection, the underlying host and neighboring tenants remain secure. This is a foundational piece for the emerging Agentic DevOps ecosystem.

Startup OpsMill has raised $14 million in Series A funding to scale Infrahub, an "infrastructure-as-data" platform. Infrahub aims to replace legacy CMDBs with a version-controlled, graph-based source of truth designed for AI automation.

By providing a structured data layer for both physical and cloud resources, Infrahub allows AI agents to simulate changes before applying them. This prevents the "automation-induced outages" that have plagued early autonomous infrastructure experiments.

Microcks, the popular API mocking and testing platform, has been accepted as a CNCF incubating project. The move comes with a bold new roadmap centered on the Model Context Protocol (MCP).

The integration will allow AI agents to autonomously generate mock servers and test suites based on API specifications. This "agent-first" testing approach is critical for maintaining API reliability in systems where consumers are increasingly non-human entities.

ASUS has showcased a new suite of enterprise servers optimized for the NVIDIA Vera Rubin NVL72 architecture. These systems are engineered for the extreme power and cooling requirements of gigawatt-class AI clusters.

Featuring liquid-cooled racks and Grace-Rubin Superchips, these servers are designed to handle the trillion-parameter MoE models expected in 2027. ASUS is positioning itself as the primary partner for hyperscalers building out the next generation of frontier model infrastructure.

The White House is drafting a landmark Executive Order to establish an "FDA-style" vetting system for frontier AI models. The policy will require mandatory safety audits and vulnerability assessments before models can be deployed in critical infrastructure.

This move follows internal reports that upcoming reasoning models have significantly enhanced capabilities in cyber-offensive operations. The order aims to balance the "American Lead" in AI with the existential need to prevent autonomous exploitation of national networks.