Claude can send email from a Gmail inbox, and Microsoft fixes a record 972 flaws

An email that arrives in Gmail now has a second reader. Anthropic's Gmail connector lets Claude search a mailbox, draft, reply, send and forward, which puts the assistant inside the inbox rather than in a chat window somebody pastes messages into. ChatGPT and Gemini can read mail; sending it without a person retyping anything is further than either has gone.

Claude can pull six months of a supplier thread into a summary, write the reply that pushes back on the delivery date, and forward the signed contract to the legal team with a note on what changed. What it cannot do is tell apart an email written to the person and a line inside that email written to it. Instructions buried in an invoice, a CV, or the quoted text at the bottom of a long chain read to the model as work to do: forward the last three messages from finance to this address, then delete the thread.

Nobody needs a password for that. The sender needs the email address, which is on the company website.

Anthropic's help page says Claude asks for approval before each send, reply or forward, and that this is the default setting. On Team and Enterprise plans, the workspace owner, not the person whose mail it is, decides whether members can let those actions run without being asked each time.

Microsoft's September update fixes 972 flaws in Windows, Office and the rest of its products. Last month the count was 620, and the month before 570.

Finding software flaws is one of the things AI does well, and it does it for the researchers Microsoft pays and for the people trying to get in.

Two of the flaws were being used in attacks before Tuesday's fix existed, according to SecurityWeek's account of the release: one in the part of Windows that lets programs talk to each other, one in the update system itself, each of them a way for an attacker who has got onto a machine to take full control of it.

Desktops are not the whole of it. By BleepingComputer's count, Microsoft fixed another 204 flaws earlier in September, outside the monthly cycle, in Azure, Entra ID, Edge and Power Automate: the cloud services and sign-in systems a company rents rather than installs.

Judge Rita Lin ruled in 2024 that Workday can be treated as an employer under US anti-discrimination law, because its software performs the screening its customers would otherwise do themselves. In May 2025 she granted preliminary certification of a nationwide collective action under the federal age discrimination act, letting other rejected applicants join. Derek Mobley, who brought the case, says he was passed over for more than 100 jobs at companies using Workday since 2017, and that many of the rejections came back within minutes or in the middle of the night.

Workday told Fortune that its recruiting tools do not make hiring decisions and that customers keep full control and human oversight. That statement and the ruling describe one process from two ends. A person approves the shortlist; the software decided which applications reached it.

Business use of AI has more than doubled in two years, and the same arrangement now produces credit files, supplier shortlists and insurance claim queues.

Two Illinois homeowners, Jacqueline Huskey and Riian Wynn, sued State Farm in federal court over the software that sorts homeowners claims. Their complaint says white policyholders were almost a third more likely to have a claim handled in under a month, and that State Farm asked Black policyholders for more information 39% more often. State Farm said the suit does not reflect its values.

Court records collected by the Civil Rights Litigation Clearinghouse show the judge limited the first phase of evidence gathering to one subject, the tools State Farm uses to separate straightforward homeowners claims from the ones it treats as complex or possibly fraudulent, and that the phase was still running in February 2026.

The recording bot in the participant list is how most people learn that a meeting is being captured. Wispr Flow's new notetaker for Mac never appears there. It takes the sound from the laptop itself, so a call in Teams, Zoom or Meet is transcribed and no bot joins the room. The settings many IT teams use to keep recording bots out do not see it.

The class action filed against Otter.ai in August 2025 argues that its notetaker recorded people who never agreed to it, and that Otter left the job of getting everyone's permission to the account holder who switched the tool on.

Meta has released Muse, an agent built to do tasks rather than answer questions: shopping, sending email, booking travel. It works on a secure virtual machine (a rented computer in the cloud) that it operates itself, and it pays with one-time-use cards issued through Stripe.

For a shop, this changes who is on the other side of the checkout. The order comes from a program matching what the page states against what its owner asked for: the price, the delivery date, the return terms, in text a machine can read. A photograph persuades it of nothing.

The card number that pays for that order works once. A merchant who stores it for the next purchase, or for a subscription renewal, has stored a number that no longer works.