ChatGPT's Chrome extension can now read and control your tabs

OpenAI's browser strategy changed shape this week: the standalone browser is gone, and the assistant is being wired into the one people already have open.
OpenAI upgraded its ChatGPT Chrome extension on October 15 so the assistant can gather information from your open tabs — and, in ChatGPT Work, act on them: sending messages, filling out forms and downloading files. The company announced the change on its own channels and documented it in its ChatGPT docs, and the Chrome Web Store listing shows the extension at 7,000,000 users with an October 15 update. Setup runs through the desktop app — install the extension from Settings, then tag Chrome with an @-mention inside a chat. Reading across tabs works in Chat; the actions that change state live in Work mode, and OpenAI's docs note that availability can depend on rollout and workspace settings.
The packaging is the story. OpenAI retired its standalone Atlas browser in August, and rather than keep paying to run a browser nobody asked for, it is now shipping browser control as an extension on top of the browser most people never switched away from. That is the same move Google made with Gemini's auto-browse, and it is the cheaper path to the same place: an agent that can see and click through the web.
It is also a new input channel into an agent with hands. OpenAI's documentation tells users to treat page content as untrusted context, and lists browsing history as elevated risk with no always-allow option — sensible defaults, and an admission that a tab full of hostile text is now something a model will read and act on. An extension that can fill a form and send a message inherits every prompt-injection problem the browser-agent category has been arguing about for a year, and it inherits them inside 7 million existing installs.
Thoma Bravo-owned Proofpoint agreed to buy data security company Varonis for about $6.3 billion in cash — $49.75 a share, a 52% premium to Varonis' price before takeover rumors surfaced in June. Reuters puts the $6.3 billion as enterprise value with equity nearer $5.71 billion; by the companies' own math that is roughly 8.4 times trailing revenue and about 53 times free cash flow. Varonis' board approved unanimously, and the deal is expected to close in the first half of 2027 pending a shareholder vote and regulators, at which point Varonis leaves the Nasdaq.
Both chief executives lead their statements with AI: Proofpoint's Sumit Dhawan frames the combination around AI agents reaching into sensitive enterprise data, and Varonis brands itself the data and AI security leader. Read that as a pitch as much as a thesis — the assets being bought are data access governance and loss prevention, not model or agent-runtime technology. But the consolidation behind the framing is real. Proofpoint bought Acuvity in February specifically for AI security across the agentic workspace, and the direction of travel is a security stack that assumes software, not just staff, is opening the files.
Ramp raised $1.85 billion at a $60 billion valuation, months after a $750 million round valued it at $44 billion, with annualized revenue now past $2 billion across more than 85,000 customers. The AI content is what makes it more than a fintech headline. Ramp runs an LLM router that sends each request to the cheapest model that can handle it, uses agents to process supplier bids and to approve expense submissions, and launched Glass this week — a tool for sharing chatbot prompts and querying business records in natural language.
A spend-management company selling model routing as a product line says something about where enterprise AI budgets are going: the pitch is no longer a better chatbot, it is control over what the chatbot costs.
What to watch: whether OpenAI publishes any abuse or incident data for tab control, and whether Varonis shareholders balk at a deal that prices the stock below where it traded in September.
Your browser tabs are now an agent's input — would you hand that over for the convenience, or is the prompt-injection surface too large? Tell us in the comments.



