CodeMingle AI News Report - September 28, 2026
Executive Summary
This Monday's briefing looks at the controls around AI-assisted work. Microsoft has outlined a broader Copilot experience, including a persistent agent that is still in preview. AWS has published agent skills for messaging tasks, while Amazon Transcribe now offers customer-managed encryption keys for custom resources. OpenAI has added an account security history view to ChatGPT.
The common question is practical: when an assistant can act through more tools, how can a person see what is available, what has changed and who is accountable for the result?
Editorial date: September 28, 2026 — Pacific/Auckland. The four source updates below are dated September 25. This is a briefing on recent developments, not a claim that they launched today. Availability and rollout details come from the linked providers; CodeMingle analysis is labeled separately.
Listen to the podcast edition
Top AI News Stories
Microsoft outlines Copilot Home, Code and Autopilot
Microsoft introduced three capabilities for its Copilot app. Home brings Chat and Cowork together and includes Office in Copilot. Code is intended to let people build solutions using technology related to GitHub Copilot. Autopilot is described as a persistent, proactive agent. Microsoft says Home and Code will begin rolling out in its Frontier program in the coming weeks; Autopilot is expanding to private preview at the end of September. These are different availability stages, not a general release of all three capabilities. Source: Microsoft, September 25
CodeMingle analysis: A persistent agent needs a visible record of its instructions, tool access, actions and handoffs. Before evaluating one, choose a bounded workflow and decide which steps require review. Assess the experience against the advertised rollout stage rather than planning as though every capability is already available to your team.
ChatGPT adds a security history view
OpenAI's September 25 release note introduces Security history in ChatGPT. It shows recent sign-ins, sign-outs and changes to multi-factor authentication, passkeys and other security settings, with time, location and device information where available. OpenAI says some details may be approximate or unavailable. On the web, the view is under Settings, then Security and login. Source: OpenAI release notes, September 25
CodeMingle analysis: This gives account holders a concrete place to investigate unfamiliar activity. A review routine should distinguish an unexpected event from an approximate location, then use the provider's account security guidance for follow-up. The history is evidence for investigation; it does not by itself establish whether an action was authorized.
Developer Tools & AI Agents
AWS publishes messaging skills for coding agents
AWS End User Messaging and Amazon SES now provide AI agent skills for the AWS MCP Server. AWS says the skills guide tasks such as verifying sending identities, building branded RCS agents and sending production email. Its announcement lists Claude Code, Codex, Cursor and Kiro among supported coding-agent environments, and says developers must connect the AWS MCP Server and add the relevant channel skill. Source: AWS, September 25
CodeMingle analysis: A natural-language request can now reach a real communications workflow. In a trial, check the exact account and recipient, the permissions granted to the agent, and the final message before sending. Keep a clear record of the request and result so the person responsible can tell whether the intended communication happened.
Data & Security
Amazon Transcribe adds customer-managed keys for custom resources
Amazon Transcribe now supports a customer-managed AWS KMS key for custom vocabularies, vocabulary filters and language models stored at rest. AWS says existing behavior continues with an AWS-owned key if customers do not opt in. With a customer-managed key, customers control key permissions and can inspect key use in CloudTrail; the announcement says the capability is available in all regions where Transcribe is offered. Source: AWS, September 25
CodeMingle analysis: For teams with key ownership requirements, test a custom resource in a non-production account before changing a production workflow. Confirm that the service can still use the key, that expected use appears in audit records, and that disabling or rotating the key has the effect your recovery plan expects. The update concerns these custom resources specifically; it does not describe every artifact Transcribe handles.
Technical Deep Dives (Architecture & Implementation)
Trace an agent's action from request to outcome
For a small messaging pilot, write down the user's request, the selected AWS account and channel, the permissions available to the agent, the proposed recipient and content, and the result returned by the service. Have a reviewer compare the result with the original request. Repeat with an ambiguous recipient and an unavailable channel. This is an evaluation exercise suggested by CodeMingle, not a description of the AWS skills' built-in controls.
Check the audit trail before depending on it
Create a custom Transcribe resource with a test customer-managed key, then confirm the expected access and CloudTrail events. Separately, review which ChatGPT account events appear in Security history and which fields may be approximate. These trails answer different questions: one concerns use of a cryptographic key, and the other concerns account activity. Neither substitutes for an application-level record of what an agent was asked to do.
Detailed Trend Analysis
The week's updates show AI workflows expanding from answering questions into creating artifacts, sending communications and continuing work over time. Provider announcements also show that visibility is being added at different layers: account security events, key-use logs and product-level action records. Our reading is that teams should connect those layers to the decisions people actually make, while checking each product's current scope and rollout stage.
Future Outlook
Watch the Frontier rollout of Copilot Home and Code and the planned expansion of Autopilot's private preview. For tools already available, a useful next step is a small, reversible trial: test one messaging task or one Transcribe custom resource, record the expected result, and compare it with the observed audit trail.