ChatGPT Work agents can now sign in to websites, without seeing the password
Po / Money and Products desk
The limit on what an AI agent could do for you was never really reasoning. It was the login wall. OpenAI has now built a way around it that keeps the password out of the model.
How the handover works
On August 25, 2026, OpenAI enabled login-required sites for ChatGPT Work on both web and mobile. The agent runs its task in a cloud browser until it hits a sign-in page. At that point it stops and hands the browser to the user. You type the username and password yourself. Then you give control back and the agent carries on.
For the duration of the takeover, screenshots are switched off. Nothing on that screen is captured, so the credentials never reach the model or OpenAI’s logs.
Cookies persist between sessions. That means one sign-in per site rather than one per task, and the login steps are not counted against message limits. The feature runs in agent mode only, on Pro, Plus, Business, Enterprise and Edu plans. The ordinary cloud browser still cannot sign in anywhere.
What actually changes
Most agent demonstrations have run on the public web, where the useful work is thin. Booking systems, invoicing portals, CRMs, internal dashboards, airline accounts and bank statements all sit behind a login. This is the first mainstream consumer agent that operates inside them.
That is also where the risk moves. An agent acting on the public web can waste time. An agent inside a logged-in account can send, buy, cancel or delete. The credential design is careful, but it addresses password theft, not the consequences of an authenticated agent taking a wrong action in a real account.
The practical question
The sensible framing is not whether the password is safe, but which accounts you would let an agent act inside at all. Read-only research in a work dashboard is one decision. Anything that spends money or sends messages on your behalf is another.
Sources
ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.
Related
Carrier Pidge delivers texts at pigeon speed, and 0.2% of them die on the way
Carrier Pidge is a free iOS app that deliberately slows down messaging by delivering each text via a simulated homing pigeon. It takes the GPS coordinates of both users, computes the real distance and flies the message at roughly 110 miles per hour, so Los Angeles to New York takes about 22 hours, with speed varying by around 25%. There is a 0.2% chance per message that the pigeon is lost to predators, hunters or distraction and never arrives. Users can track flights live, manage a flock and acquire more birds. Developer Noah Iarrobino built it in about three weeks and released it on the App Store on April 19, 2026; it reached version 1.7 on August 22 and holds a 4.6 rating from 527 ratings. There is no Android version.
Claude has watermarked its text for three weeks and there is still no detector
Every Claude model launched on or after August 2, 2026 embeds a statistical watermark in the text it generates, with older models to follow over the coming months. Anthropic's own explainer says it will "soon" offer a watermark detection API and that it is still working out the implementation details, so as of late August there is no public way to check a passage. The method is SynthID-Text, published by Google DeepMind and tracing back to a 2022 proposal by Scott Aaronson: a secret key plus the preceding words replaces ordinary randomness in the model's word choice, with nothing added to the text and no hidden characters. Longer passages are easier to confirm, light editing survives while a full rewrite does not, and factual writing and code carry the sparsest marking because they allow the fewest phrasing choices.
Chinese state hackers more than doubled their attacks after adopting DeepSeek
Taiwanese research firm TeamT5 told Bloomberg on August 24, 2026 that state-affiliated Chinese hacking groups have more than doubled the number of attacks they carry out since delegating routine tasks to open-source AI models and using them to develop malicious software. DeepSeek is the model of choice; TeamT5 chief analyst Charles Li attributes that to it being relatively powerful with very low cyber guardrails. The report names specific uses: Grimfengxi generated exploit code, Huapi targeted a Taiwanese company's email system, and Teleboyi collected 1,000 IP addresses and mapped a target's domains. Researchers say they have not yet observed a more expensive model such as Kimi K3 used in an attack. The UK AI Security Institute warned in May 2026 that models' cyber capabilities are doubling every few months.