Last Week in AI: 14–20 September 2026

The week of 14–20 September 2026

This week’s selection is for teams responsible for business systems, including those commissioning software from suppliers. A code-security tool received an operational update, a cloud security service opened a preview route for AI agents, and Anthropic announced a different approach to external evaluation.

We have checked the original release notes and announcement, not tested these products. The distinctions matter: an updated preview tool, a new preview integration and an evaluation partnership are different kinds of development. None should be read as a recommendation to grant an assistant unrestricted access.

CodeMender changes what runs in its CLI

Google’s 16 September release notes list CodeMender v0.8.0, with Gemini 3.8 Flash becoming its default model for CLI sessions. The update also addresses Windows command execution when Git Bash is installed outside expected locations, and problems retaining verification results between sessions.

The access conditions deserve equal attention. Google describes CodeMender as a limited-customer Public Preview, obtained through its sales team, for supervised testing and evaluation. Its documentation explicitly excludes commercial or production use under those preview terms. This is an update to a preview tool, not a general production release.

For an IT manager working with a development supplier, our suggested question is which model and tool version produced the security evidence being presented. If you already have authorised evaluation access, use a permitted test repository and compare the recorded findings before and after an upgrade. Ask whether a reviewer can reproduce the result; do not equate a model change with independently demonstrated improvement.

Security Command Center gets an agent interface in preview

On 16 September, Google added preview MCP endpoints for investigative and management tasks in Security Command Center. MCP is the connection mechanism named in the release; the announcement establishes that this interface is available in preview, not that an agent’s conclusions or changes can be trusted without checking.

The practical opportunity for an existing Google Cloud security team is to investigate whether an assistant can help with a tightly defined piece of its workflow. Our suggested starting point is a question about findings that the analyst can independently answer. Keep any proposed configuration change subject to the team’s normal approval.

Before arranging a trial, ask the service owner to confirm applicable access, permissions and preview terms. Set a boundary between retrieving information and changing the environment. This is our evaluation advice, not a claim that we have verified your tenant configuration or the connector’s safety.

Anthropic plans embedded evaluation with Accenture

Anthropic announced its Accenture partnership on 18 September. Faculty, Accenture’s specialist AI business, will lead work including model evaluation, red-teaming and safeguard testing. Anthropic describes evaluators working inside AI companies with access comparable to employees, while noting that operating details are still being developed.

This is an announcement about evaluation capacity, not a published verdict that a particular model is safe. For a business buyer, it raises a useful question for assurance conversations: what evidence can the evaluator actually inspect, and what results will you be allowed to see?

Our recommendation is to ask suppliers to separate their testing plans from completed findings. Request the model version, scope, limitations and date of any evaluation they rely on. A partnership announcement can be relevant context, but it does not settle whether a system is suitable for your organisation’s work.

A useful question for your next supplier call

Ask for one piece of evidence that links an AI feature to the workflow you intend to use, then ask what would invalidate that evidence after a model, tool or permission change. Record the answer alongside the proposed trial boundary and its owner.

We have left out the Salesforce-in-Claude beta covered in our separate Instagram post this weekend. These are distinct developments from the previous roundup’s agent sandboxes and coding-model launch. Which would be worth examining more closely for your team: reproducible code-security findings, an assistant for cloud investigation, or stronger evidence behind supplier assurances?