Daily AI · September 28, 2026
Useful AI Daily - September 28, 2026
The practical AI question this Sunday is not which agent can do the most. It is whether access follows the right person, every extension has a review path, and a failure leaves evidence someone else can inspect.
This is a weekend operating brief built from September 24–25 releases and one still-useful September safety baseline. Try one connector or plugin with a narrow scope, keep the originating conversation attached to the work, and make review the default for data that would matter if it appeared in the wrong channel.
The Short Version
- A plugin directory is also a distribution and trust surface.
- Personal access should follow the requester, not silently become channel access.
- More chat context is useful only when the resulting work keeps a trace back to it.
- Security automation should enrich evidence, not turn a weak signal into an automatic fix.
- A disclosed failure needs a date, scope, impact and an accountable mitigation path.
5 Updates Worth Your Time
Treat a plugin submission as a product review, not a listing form
- What changed
- Claude introduced a developer portal for submitting plugins to its directory. It supports a single remote MCP connector or a GitHub-hosted bundle of MCP servers and skills, with submission checks, safety scanning, review feedback and post-launch usage analytics.
- Why it matters
- For an indie builder, distribution now comes with a visible review path and a reason to define the plugin's data scope before users install it.
- Try, watch, or skip?
- Before submitting any extension, write a one-page card: inputs, outbound calls, stored data, permissions, owner, support contact and a revoke path. Use the portal's scan as one gate, not as proof that the integration is safe.
Keep personal connectors personal inside team channels
- What changed
- Claude Tag can now use a requester's own connected calendar, drive, CRM or other tools in a Slack channel. The announcement says others cannot use those connectors, and users can review a response before it posts; enterprise administrators can require review.
- Why it matters
- This is the ordinary-user bridge: a shared chat can help you compare a private document with team work without making that document a shared channel resource.
- Try, watch, or skip?
- Start with a calendar lookup or a non-sensitive project note. Confirm whose credentials are used, who can see the result, whether posting is manual, and how the connector is removed.
Make shared-chat work point back to its source
- What changed
- GitHub says Copilot can use more supported Slack and Teams context, check for similar issues, create GitHub work with links back to the originating discussion, and offer model selection for a conversation. The capabilities are a gradual public preview for eligible plans.
- Why it matters
- Context can reduce repetitive explanation, but it becomes hard to audit when a ticket or pull request loses the conversation that prompted it.
- Try, watch, or skip?
- For one low-risk issue, require the generated work item to include the source thread, a short human-written problem statement and a named repository owner before anyone starts implementation.
Use AI to enrich security records before it changes anything
- What changed
- NIST updated materials for its AI-agent workflow for enriching National Vulnerability Database information. Its stated goal is to improve the scalability, automation, interoperability, transparency and utility of vulnerability records as AI-assisted discovery increases.
- Why it matters
- The useful pattern is not 'let an agent close alerts.' It is 'use an agent to organize evidence so a security owner can validate priority and remediation.'
- Try, watch, or skip?
- Have an agent summarize one public vulnerability record into affected asset, evidence links, exploit preconditions, owner and proposed next step. Do not let it alter production tickets, packages or controls without human review.
Make incident disclosures usable by someone outside the lab
- What changed
- As a still-practical September 16 signal, OpenAI published a framework and six reports on unexpected or concerning model behavior. It says reports should describe the behavior, severity, context, dates, discovery and known or planned responses when available.
- Why it matters
- For builders and buyers, a vague assurance is less useful than a record that lets another person judge scope, uncertainty, mitigations and whether a similar workflow should be paused.
- Try, watch, or skip?
- Add a lightweight incident template to your AI product now: what happened, data involved, who was affected, when it was detected, immediate containment, next verification and the owner of the follow-up.
Tool Worth Trying Today
A 20-minute connector and approval rehearsal
In a shared channel, ask an assistant to compare one non-sensitive personal source with one public team source. Review the draft before posting, verify that nobody else can invoke the personal connector, then revoke access and confirm the workflow stops.
Best for: Small teams testing connected AI without turning a shared channel into a shared credential store.
Watch out: Use public or synthetic material for the rehearsal. Do not upload sensitive personal, customer, legal, unreleased or private photo/document data before retention, terms and administrator controls are clear.
Privacy / Cost Watch
- A channel integration can expose personal connectors, files or thread history in the wrong context. Check the identity, scope, review setting and revoke control before the first real task.
- Do not treat a directory safety scan, a connector approval screen or an agent summary as a substitute for your own data classification and human review.
- Price for connector calls, retrieval and long-running agent work separately from the model prompt; usage can grow after an integration becomes part of a recurring workflow.
- For security, health, legal or financial decisions, use official records and qualified review rather than relying on an agent's confidence or condensed explanation.
One Practical Workflow
Run a 30-minute permission and evidence pass
- Choose one connector, plugin or chat-to-ticket workflow and replace private inputs with a synthetic sample.
- Write down the requester identity, accessible systems, allowed output locations and every action that needs approval.
- Run the workflow once and verify the result links back to its source conversation or evidence record.
- Review the output before it posts or changes another system; record the owner, spend limit and rollback route.
- Revoke the connector or disable the plugin, then confirm that new requests cannot keep using the old access.
Builder Note
A useful agent can carry context from chat to ticket, plugin to tool, or connector to answer—but authority should not broaden just because the surface changes. Design for a visible requester identity, explicit scope, reviewable output, usage limits and a one-click revoke path.
Ignore For Now
A connector catalogue without boundaries
More integrations are not automatically more useful. Skip any extension whose data destinations, posting behavior, cost model or off switch you cannot explain in a short sentence to the person whose data it touches.
Bottom Line
September 28 is a permission-and-proof review. Recent connector and plugin releases make it easier to move work between people and systems; the durable advantage is making every handoff attributable, reviewable, affordable and reversible.
Sources
- Claude: Build plugins for Claude with the directory submission portal
- Claude: Claude Tag now supports personal connectors in channels
- GitHub Changelog: Updates to GitHub Copilot for Slack and Microsoft Teams
- NIST: Development of an AI Agent Enrichment Workflow at the National Vulnerability Database
- OpenAI: Our framework for reporting model misalignment