ChatGPT and Claude now learn a work task from one screen recording

Teaching Claude a work task now starts with a screen recording. Anthropic added Record a Skill to Claude Cowork on 21 July: a person does the job once, talking through the reasoning while the screen is captured, and Claude turns that into a skill it can run again. The option is in the + menu of the Claude desktop app, for Pro, Max and Team subscribers. OpenAI shipped the same idea on 18 June, as Record & Replay in the Codex Mac app, and named filing an expense report as one example.

Describing a routine job in words is the part that keeps failing. Nobody writes down that the invoice with the split delivery goes to the second approver, because nobody thinks of that as a step. On a recording it simply happens.

Neither tool retrains its model. OpenAI's documentation says the recorded skill is written out as text: when to use the workflow, what inputs it needs, the steps, and how to check the result, so a person can read it and correct it. The automation software many finance teams license today replays saved click positions and stops working when a button moves; steps written in plain words still apply after the screen changes. What a recording cannot supply is any way of noticing that the answer was wrong.

Anthropic's feature captures screen activity, clicks, keystrokes and voice narration together, so anything else open behind the task goes into the recording too. Codex's version runs on macOS only, and at launch OpenAI excluded the European Economic Area, the UK and Switzerland.

The NSA, CISA, the FBI, the Department of Energy and the Environmental Protection Agency issued a joint advisory on Wednesday about Siemens S7 series programmable logic controllers, the small industrial computers that operate pumps, valves and production lines. Attackers are using AI to write exploitation scripts and disguising them as ordinary monitoring software, the agencies say, which gives them read and write access to the controller's memory and to the program that drives the equipment. The agencies wrote that this is "not a theoretical risk", and told operators to act with urgency.

None of this required a new flaw in Siemens equipment. The controllers being reached are the ones facing the internet, found through public scanning services such as Censys and ZoomEye, running outdated software or default passwords. Writing working attack code for an industrial protocol used to take specialist knowledge that few intruders had, and the AI supplies that part now. The sectors listed as most targeted are critical manufacturing, energy, water, chemical plants and food production.

On 26 and 27 July, a coordinated attack disrupted more than 30 community water systems in Minnesota. In Braham, a city of about 1,700 people, it shut down the operating controls and with them the well and the treatment plant; public works crews had it running again in about two hours.

The advisory describes what the attack traffic looks like on a network: repeated connection attempts on port 102, and write operations outside the scheduled change windows.

Calendly began switching on two AI features on 19 August: Notetaker, which joins a call, records and transcribes it and drafts the follow-up email, and Callie, a scheduling assistant it is still testing. Both work across Zoom, Google Meet and Microsoft Teams. Because the notetaker is joined to the scheduling system, it can tell attendees the call will be recorded before the meeting starts.

In the consolidated federal case over Otter.ai's notetaker, Judge Eumi Lee heard the motion to dismiss on 20 May and has not ruled on whether a recording bot in a video call is interception under the federal Wiretap Act. A February analysis by the law firm Duane Morris said tools that record without the affirmative consent of everyone present may expose the employers and individuals using them, as well as the vendor.

The most-quoted figure of the year on business AI comes from a 2025 MIT study: 95% of custom enterprise AI projects return nothing financially, while general, ready-made tools succeed far more often.

A general tool goes into the work as it is done today and helps one person finish their own step faster. A custom build usually automates a single step inside a chain that still carries all the same approvals around it, so the invoice waits as long as it did before and the finance director sees no change in the month-end numbers.

That is the case an IBM executive is now making: the return arrives when a whole workflow across teams is automated from end to end. IBM sells the consulting work that redesigning those handoffs takes.