Mistral released Shieldstral on August 4, a 3.8B-parameter open-weight multimodal moderation model that runs on a single 16GB GPU and lets operators define policies in plain English at inference time.

• Performance vs. size: Shieldstral achieves 84.9% average F1 on text safety benchmarks, matching GPT-OSS-Safeguard-20B despite being roughly 5x smaller, and beats competitors up to 4x its size on image safety.
• Policy flexibility: Unlike Llama Guard or OpenAI's Moderation API, which use fixed taxonomies, Shieldstral converts any natural-language policy into a calibrated yes/no probability in a single forward pass — no fine-tuning required to change rules.

Microsoft launched Windows 365 for Agents, a product that provisions Entra ID-joined, Intune-managed Cloud PCs specifically for AI agents to operate legacy apps and portals that lack APIs.

• Governance-first design: Each agent Cloud PC carries full audit trails and runs under existing Intune policy, meaning security teams govern agent compute the same way they govern user compute — no separate control plane.
• Legacy system unlock: The product targets the large installed base of on-prem and browser-based apps that will never get API-first agent connectors, making it a direct complement to Copilot Studio and Azure AI Foundry agent deployments.

Delinea released runtime authorization capabilities that enforce identity policy on specific agent actions inside an active session before they execute, not just at connection time.

• Gap it fills: Existing PAM and identity tools verify the session open and audit after the fact; Delinea's approach intercepts individual commands on databases, Kubernetes clusters, cloud consoles, and MCP servers while the session is live.
• Agent credential risk: Autonomous agents routinely inherit broad credentials from the orchestrating human account — this capability is designed to prevent agents from executing actions outside their sanctioned scope even with valid credentials in hand.

Cursor Research released Mixture-of-Kittens (MoK), an open-source training kernel for mixture-of-experts models that fuses communication and computation into a single GPU kernel, reporting 41% overall throughput improvement on 512 Nvidia GB300 GPUs.

• Benchmark claim: MoK records up to 2.37x faster performance than public alternatives in single MoE layer tests by eliminating the sequential communication-then-compute pipeline that creates bottlenecks at scale.
• Why it matters beyond Cursor: Publishing the kernel under an open license means any lab training MoE models on large NVL72 clusters can adopt it — a meaningful infrastructure contribution that also signals Cursor is investing in vertical integration across training and inference.

The U.S. Army Human Resources Command became the first Department of War organization to deploy Salesforce Agentforce 360 under a new IL5 authorization, providing 24/7 AI agent support to 9.2 million soldiers, veterans, civilian staff, and military families.

• Scope: The deployment covers defense logistics, recruit onboarding and training, administrative support for military families, and real-time command data insights — all operating on controlled unclassified information under Impact Level 5 security requirements.
• Platform signal: Slack accounted for nearly half of Salesforce deals over $1M ACV in Q1 FY2027, with bookings in that tier up 80% year over year, suggesting Agentforce's federal expansion is riding a broader enterprise momentum rather than being a one-off government win.