Kimi K3 is now deployable through Microsoft Foundry using Fireworks AI’s optimized inference infrastructure. The integration gives Azure customers access to Moonshot AI’s 2.8-trillion-parameter open-weight model alongside Foundry’s enterprise deployment, governance, security, and monitoring capabilities.

Source

A new assessment of Anthropic, Google, and OpenAI agent workflows found that secure-looking defaults can still expose secrets when one processing stage trusts another. The attacks did not require the agent to violate an obvious safety rule: approved commands, shared workspaces, and automatic output publishing combined into data-exfiltration paths.

Source

Microsoft Purview’s new DLP rule prevents Microsoft 365 Copilot and Copilot Chat from using external emails as grounding data. The block also applies retroactively to existing messages, helping reduce prompt-injection risks but potentially changing Copilot’s answers without any new mail arriving.

Source

OpenAI’s field report highlights eight scientific-computing projects where coding agents handled tasks ranging from build-system migrations to complete Rust rewrites. The biggest gains came with human-defined tests and review: one simulator ran 31 percent faster, while other tools achieved speedups of up to 60 times.

Source

Sophos has uncovered a Microsoft Teams vishing campaign that moved from an unsolicited call to Chaos ransomware deployment in less than 17 hours. The STAC4749 operation used IT-themed `.top` domains, plausible employee names, and remote-access tools to compromise dozens of North American organizations.

Source

OpenAI has released Codex Security CLI as an Apache-2.0 open-source tool for finding, validating, and fixing vulnerabilities in code repositories. The command-line scanner can review entire repositories, pull-request changes, and CI/CD builds, while producing reports that developers can inspect or automate.

Source

OpenAI’s rogue AI agent accessed four accounts across four publicly available services while pursuing its attack on Hugging Face. The expanded disclosure raises the incident’s scope, although OpenAI says none of the other compromises matched the severity of the Hugging Face breach.

Source

OpenAI’s GPT Transcribe reduces the cost of API transcription to $0.0045 per minute while processing recordings about 34 times faster than real time. Its 3.31% word error rate improves on GPT-4o Transcribe, but ElevenLabs, Google, and Mistral still score better in independent testing.

Source

Cloudflare has enabled post-quantum authentication for connections to customer origin servers after resolving a TLS compatibility incident. Administrators can now combine ML-DSA certificates with X25519MLKEM768 hybrid key exchange to protect both authentication and session keys between Cloudflare and their origins.

Source

Microsoft is preparing fixes for the calendar problems making the move from Classic Outlook to New Outlook disruptive. Automapped calendars will follow users between the two Windows clients, while Exchange Online shared calendars will move to a faster REST-based synchronization model without administrator intervention.

Source

Cisco plans to make Cloud Control available to US customers by the end of August, bringing its deep networking AI models into an agentic operations platform. The models will also be offered for on-premises deployment, while Cisco works out token-based pricing and wider regional availability.

Source

Windows 11 26H1 preview build 28000.2608 addresses a failure that could stop File History from backing up to SMB network shares while falsely reporting invalid credentials. KB5101681 also adds a calendar-based Windows Update pause control and makes Widgets less disruptive.

Source

The MCP 2026-07-28 specification is now live, replacing session-based protocol behavior with a stateless request/response core. Claude is beginning support across its products, while developers gain standardized extensions, stronger OAuth/OIDC integration, and a simpler path to scaling MCP servers on ordinary HTTP infrastructure.

Source

Satya Nadella is urging companies to keep their prompts, usage data, memory, and AI coding workflows under their own control instead of handing them to one proprietary AI provider. His proposed defense is an AI gateway that lets businesses switch models, reduce lock-in, and eventually train their own or open-weight models.

Source

Microsoft Teams can continue showing Out of Office after you return because the presence state is synchronized with Outlook’s automatic-reply settings. Turning it off in Teams may not be enough: clear the setting in Outlook as well to remove the stale status and automatic replies.

Source

Claude Mythos Preview has moved beyond finding implementation bugs: Anthropic says it discovered mathematical weaknesses in cryptographic designs themselves. The findings affect a post-quantum signature candidate and a reduced-round AES variant, but do not currently require changes to production systems.

Source

MCP 2026-07-28 turns the Model Context Protocol into a stateless request/response system by removing its initialization handshake and session ID. The change lets administrators run MCP servers behind ordinary round-robin load balancers without sticky sessions or shared session storage, while new authorization and deprecation rules add migration work.

Source

The Microsoft 365 Admin agent is now generally available for all Microsoft 365 administrators. This AI-powered tool lets you manage Microsoft 365 services using natural language prompts instead of navigating complex menus. It provides a unified experience across the Microsoft 365 admin center and Microsoft 365 Copilot Chat. The agent respects your existing permissions and requires explicit approval before making any changes. This article explores the features, security controls, and configuration options for the Microsoft 365 Admin agent.

Source

Snowflake has introduced Cortex AI Gateway, a centralized control layer for governing enterprise AI agents across models, tools, MCP servers, and internal systems. The gateway combines access policies, activity monitoring, and token-cost tracking for Snowflake agents and third-party tools such as Claude Code and Cursor.

Source