Data & AI watch, week of 19 to 25 October 2026

Announced on 20 October: the integration that connected a Microsoft Foundry agent to a Fabric data agent through a custom tool is supported until 6 November 2026. Its replacement is the Fabric IQ (OneLake Catalog) tool, in preview, which exposes each published data agent as an MCP tool. A single Foundry agent can now call several data agents, picked by name from the catalog. The steps: publish the data agent with a detailed description, add the Fabric IQ tool in Foundry, update the instructions, remove the old tool, then test.

Advanced SQL and DAX query generation in Fabric data agents is now generally available. Before writing its query, the agent explores the schema, tests joins and looks up the values actually stored, so it maps "California" to the stored value "CA". On semantic models, answers are consistent across the agent, Power BI and Microsoft 365 Copilot. The improvements reach the standard runtime from 2 to 10 November. Two settings arrive for SQL sources, in preview: schema descriptions, and Markdown "topics" consulted only when the question relates to them. On its own tests, Microsoft reports DAX accuracy rising from 89% to 92% and median response time almost halved.

Four language additions arrive in preview: GROUP BY ALL and ORDER BY ALL, which reuse the SELECT columns, QUALIFY, which filters after a window function without a subquery, and the FROM … SELECT form, which starts with the source. The warehouse and the lakehouse SQL analytics endpoint gain MEDIAN, QUANTILE and their approximate versions, plus aggregate syntax for PERCENTILE_CONT and PERCENTILE_DISC. Finally, a T-SQL external function can call a Python function from Fabric User Data Functions, from a SELECT, INSERT or UPDATE: sending a Teams message, reading a secret from Azure Key Vault, checking an IBAN or a VAT number.

The migration assistant for Fabric Data Warehouse now accepts Teradata as a source (preview): a ZIP archive of SQL and BTEQ scripts is translated to T-SQL, tables, views, procedures, functions and macros are created, and whatever needs rework is grouped in an "Objects to fix" list. The assistant also handles incremental migration: one business domain at a time, into an empty, previously migrated or existing warehouse, with naming conflicts flagged and direct dependencies added automatically. It migrates definitions; data is loaded in the next step.

In Fabric real-time dashboards, you describe the operational view you need in plain language, and Copilot generates the KQL query and an HTML visual, added as a new tile. The same day, Eventstreams learned to tell apart several event types within one stream, each with its own schema, and to process each one separately.

On 21 October the PHP project released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, fixing two vulnerabilities rated 9.2 out of 10 (CVE-2026-109618 and CVE-2026-109619). They allow unauthenticated remote code execution, with the highest risk for mod_php. Minimal backports exist for the 7.4 to 8.1 branches, and updating remains the only complete protection. The security firms Calif, working with OpenAI, and Wiz found them with the help of AI models. As a precaution, Finland took some public online services offline and the BBC some of its pages.

The EN 18286:2026 standard, developed by CEN and CENELEC, has been published in the Official Journal of the EU. It describes the quality management system that Article 17 of the AI Act requires from providers of high-risk AI systems, and applying it gives a presumption of conformity with those requirements. It is organised around the AI system lifecycle and layers on top of ISO 9001, ISO/IEC 42001 or existing sector standards. The Commission expects further harmonised standards to follow.

As announced in April, OpenAI retired a series of older models from its API on 23 October, including gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, gpt-3.5-turbo, o1, o3-mini, o4-mini and gpt-image-1. Fine-tuned models follow their base model. OpenAI recommends gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna as replacements, depending on the case. Next dates: Evals become read-only on 31 October, and the Evals platform, the v1/prompts API and Agent Builder shut down on 30 November.

Anthropic is handing the Agent Skills open standard, which describes an agent's skills and working methods, to the Linux Foundation's Agentic AI Foundation, already home to the Model Context Protocol. Engineers from Google and OpenAI will maintain it alongside Anthropic's. According to Anthropic, more than 70 platforms have integrated it, including ChatGPT, GitHub Copilot and VS Code. Versioned releases, a formal proposal process and security guidance for installing skills are planned.

OpenAI and Google Cloud announced that OpenAI models and Codex agentic workflows are coming to Model Garden, with Google Cloud's access controls and data governance. According to OpenAI, purchases will go through Google Cloud Marketplace and can count towards existing spending commitments, with general availability planned in the coming weeks.