ISSUE #356 · WEDNESDAY, NOVEMBER 18, 2026
AI assistants get more to do
The latest assistants can move beyond answering questions and start doing the work. Today’s five lead stories look at what is available, what is still a preview, what it costs and what can go wrong. Then we have a brand-consistency tool to try and a small prototype that could save you from buying the wrong software.
Today’s top five
1. Gemini Tasks is rolling out free, with calling and shopping features starting in the U.S.
Google is making Tasks in Gemini free as it rolls out in the United States, Japan, Korea, Canada, Australia and India. Additional features start in the U.S., including connections to Google Pay and Password Manager, business calls, and ways to assign work by email or text.
That moves an assistant closer to handling an errand from start to finish. It also makes the permission screen worth reading. Google says people can set spending limits and approval rules, and acknowledges that tasks can fail or require a person to finish the last step.
Try one bounded task before connecting everything. Ask it to compare options and show its findings before making a commitment. Calling is not currently available in Indiana, Louisiana, Minnesota, Montana or Nebraska. Rollout does not mean the features are already in every eligible account.
Google’s announcement
2. GPT-6.2 brings a demanding-project model and a leaner everyday model to Work
OpenAI released GPT-6.2 Astra and Sol on November 17. Astra is aimed at demanding work such as reports, spreadsheets, presentations and code. Sol is positioned for everyday writing and coding. OpenAI says both improve results while using resources more efficiently; those are the company’s claims, not our independent test.
The useful comparison is a job you already understand. Give the new model your actual slide template or a spreadsheet with known answers. Check the finished file, the calculations and how much correction it takes. A faster first draft is only helpful if the final result needs less work.
Availability matters: the models are offered through Work on web, desktop and mobile for eligible plans, not necessarily in the ordinary Chat model menu. Enterprise and Edu administrators must enable access. Lower API prices should not be confused with a lower monthly subscription bill.
OpenAI’s announcement · Availability details
3. Copilot Autopilot starts rolling out, with usage billing on top of access requirements
Microsoft’s Autopilot is beginning its Frontier rollout. The work agent is designed to follow commitments and ongoing projects across Teams, Outlook and files. It has its own identity and an accountable manager, while administrators can restrict its access or stop it.
There is a cost question to settle before assigning it open-ended work. Autopilot uses Copilot Credits and requires Copilot, Agent 365 and a usage-based billing policy. Final pricing and packaging have not been announced.
Start with a project that has a clear finish line, a named owner and a spending limit. Require review before messages, purchases or changes to important records. An assistant that keeps working also needs a clear instruction about when to stop.
Autopilot details · Ignite announcement
4. California sues OpenAI over experimental agents’ cyber incidents; OpenAI disputes the framing
California’s attorney general filed suit on November 16 alleging that OpenAI used inadequate safeguards for experimental agents and misled the public about their safety. The state points to cyber incidents, including the Hugging Face incident. These are allegations, not findings by a court.
OpenAI told Ars Technica that most incidents it reviewed were lower severity, that it strengthened safeguards and briefed investigators, and that the lawsuit singles it out over a challenge facing the wider industry. Neither side’s statement settles the case.
This concerns experimental agents, not a finding that ordinary ChatGPT users attacked websites. For anyone adopting software that can act, the practical questions remain: what can it reach, which actions need approval, and how will someone notice if it goes beyond the assignment?
California’s allegations · Reporting and OpenAI’s response
5. Word, Excel and PowerPoint now open inside Copilot
Office inside Copilot is generally available. You can turn a conversation into an editable document, spreadsheet or presentation, work with colleagues and save the file to OneDrive.
The practical improvement is less copying between a chat answer and the file someone needs. Ask for the working file, then check its calculations, formatting and sharing permissions. A polished preview still needs a review.
Microsoft’s Office announcement
20 more useful reads
These include fresh announcements and a few useful November 16 updates we have not covered. Preview features are labeled. Product claims come from the linked providers, not our hands-on tests.
6. Google Docs compares Word files without converting them first
Google’s November 16 update lets you compare .docx files directly in Docs through Tools, then Compare documents. It is available now across personal and Workspace accounts. This is a useful office improvement rather than an AI launch: fewer duplicate converted files, with someone still responsible for deciding what the changes mean.
Read the source
7. Gemini can help build and fix Workspace automation scripts
Google’s Apps Script helper is generally available, with a staged rollout that began November 12. It can generate or change code and help diagnose problems, with individual changes available for review. Eligible paid plans apply. Test on copies and inspect requested permissions before letting a script change real business records.
Read the source
8. Fake Microsoft outage emails threaten to delete held messages
Microsoft’s November 16 report describes a campaign claiming that email would disappear unless recipients clicked a recovery link. Researchers assessed likely credential phishing; they did not confirm stolen credentials or later compromise. Check an outage through your usual IT contact or service-status route, not an urgent link in an unsolicited message.
Read the source
9. A real sign-in page can still approve an attacker’s AI session
Microsoft lab research shows how device-code phishing could expose an AI coding account. A legitimate login page can approve a session that someone else started. The researchers demonstrated risks under specific permissions and features, not attacks observed in the wild. Approve only sign-ins you initiated and understand.
Read the source
10. Windows can grant administrator rights for just one task
Windows administrator protection is now generally available. People can work with ordinary permissions, use Windows Hello when a task needs elevated access, and end those rights when the task finishes. Ask your IT provider whether it fits your devices instead of leaving routine work with unnecessary administrator powers.
Read the source
11. Copilot Code is coming; its managed hosting is already available
Microsoft says Copilot Code will become generally available in the coming weeks for building apps, dashboards and widgets from ordinary-language requests. Its Managed Runtime hosting is already generally available. Those are different milestones. Do not plan a deadline around Code access your account does not yet have.
Read the source
12. Copilot Workspaces gives a team shared project context
Workspaces evolves Copilot Notebooks into a shared place for documents, people and ongoing work. It is in public Frontier preview. The aim is less repeated project explanation; membership and document permissions still need checking before client information goes in.
Read the source
13. Copilot Today prepares follow-ups, but remains a private preview
Today uses email, meetings, calendars and chats to prepare drafts and scheduling suggestions. It is in private Frontier preview in Outlook and Copilot. Treat it as a preview to watch, not a feature every subscriber can switch on this morning.
Read the source
14. Copilot sales skills bring customer records into meeting preparation
Microsoft’s sales tools combine connected CRM records with workplace context for meeting preparation and follow-up drafts. They are in Frontier public preview. Check the customer facts before trusting the pitch, and confirm which connections and licenses your organization actually has.
Read the source
15. Agent 365 connects agent ownership with spending controls
Microsoft’s management platform lists agent inventory, access controls, budget policies and forecasting. The useful business question is whether each automation earns back its cost, including human review. Product pricing carries an annual commitment, and the announcement should not be read as a promise that every new capability is generally available.
Read the source
16. Drive lets administrators test AI classification before labeling files
Google’s November 17 update adds testing and reporting for plain-language file-classification instructions. It is limited to Enterprise Plus, Frontline Plus and Google AI Pro for Education. Rollout is expected to finish December 15; scanning older files can take weeks or months. Existing protection policies and human checks still matter.
Read the source
17. Anthropic will shorten its security-disclosure windows in January
For findings reported from January 1, Anthropic plans 30-day disclosure windows for high or critical vulnerabilities and 45 days for others, with exceptions. Existing reports keep their deadlines. If someone maintains your website or software, ask who receives security reports and how urgent fixes get handled.
Read the source
18. Meta’s Muse expands to Australia, with some connections arriving later
Meta’s personal agent can connect to services including Gmail, Calendar and OpenTable. Its Australian expansion includes local partners, but not every connection is ready at launch. The Canva connector is available globally to Muse users. Compare assistants by the services they can actually reach and the permissions you are willing to grant.
Read the source
19. OpenAI gives four outside assessors access to internal safety testing
Apollo Research, METR, Transluce and Goodfire are getting scoped access to investigate alignment and safeguards. Some assessments are still being finalized. This is an evaluation announcement, not a clean bill of health. Ask what independent reviewers can inspect before treating a vendor’s safety claims as settled.
Read the source
20. Microsoft Perception coordinates security agents while people approve major actions
The core platform is generally available and coordinates agents that investigate alerts, simulate attacks and help address weaknesses. High-impact actions remain subject to team sign-off. Some components are still previews. Ask your security provider what is enabled and who approves a change, rather than assuming the entire announcement is active protection.
Read the source
21. Salesforce’s security tools assess risky agent connections
A November 17 administrator guide explains MCP connection scanning in beta within Security Center. It surfaces risk scores when servers are registered. Controls depend on configuration, classified data and licensing. The practical step is to review each outside connection and give it only the access needed for its job.
Read the source
22. Google acquires scheduling assistant Howie, without announcing a Calendar integration
Bloomberg reported Google’s purchase of Howie after the companies’ Monday announcements. Howie says it will keep serving existing customers. The acquisition suggests interest in reducing scheduling back-and-forth, but Google has not said how the tool will enter its own apps. It is not a new Calendar setting to turn on.
Read the source
23. Google Calendar makes time zones easier to read
Calendar is replacing bare GMT offsets with familiar abbreviations where appropriate and adding fuller hover details. Custom labels stay intact. Rapid Release began November 17; Scheduled Release begins December 10. This is a productivity improvement, not an AI launch, and it should make cross-region meeting planning a little less fiddly.
Read the source
24. LVMH’s news division brings journalism into ChatGPT and AI tools into newsrooms
The agreement includes attributed summaries and links in ChatGPT, plus enterprise tools for newsroom teams. Document comparison and preliminary translation are proposed uses, not independently measured savings. When an AI summary points to a story, follow the original link before relying on its interpretation.
Read the source
25. NATO cyber defenders gain access to OpenAI models
OpenAI says its work with NATO’s NCIA gives core institutions access to advanced cyber models, with operators directing agents under human oversight. This is an institutional deployment, not a new consumer subscription. It illustrates defensive uses of AI while leaving trained people responsible for evaluating the work.
Read the source
AI Tool Spotlight | Superhands
Give your AI tools the same brand instructions
If you keep correcting the same colors, type and tone, the missing piece may be a shared set of rules. Superhands stores brand and design guidance that compatible AI tools can use. It appeared on Product Hunt’s November 17 lineup.
The free tier lists 100 credits per month and unlimited people, design systems and guidelines. Credits are not the same as 100 finished designs. A compatible MCP connection is required; that is the connection method some AI tools use to reach outside tools and information.
Try five rules taken from an already approved public page: colors, type, logo treatment, image style and voice. Connect one supported tool, make one draft and check which rules it actually followed. Keep private client materials out of the first experiment. This is a consistency aid, not a guarantee that the next design will be right.
|