OpenAI has brought "computer use" to Chrome users, turning the world's most popular browser into an agentic workspace. In Chat mode, ChatGPT can look across your open Chrome tabs and gather information to give more useful answers. In Work mode, it goes further: it can control Chrome tabs directly — filling out online forms, sending messages, downloading files, organizing and grouping tabs, and collecting information across websites into comparison sheets. OpenAI also updated ChatGPT's voice capabilities alongside.
What each mode does
Chat mode is read-only but cross-tab: ChatGPT sees what's open and synthesizes across it — comparing products across shopping sites, explaining a long article, comparing subscription plans. The shift is context: instead of copying and pasting between tabs, the assistant perceives your browsing session as its working context.
Work mode is the agentic leap: ChatGPT takes actions inside your tabs — completing web forms, sending messages, downloading files, tidying and grouping tabs, scraping information across sites into comparison sheets. It moves from answering questions about the web to operating the web for you: research assistant to virtual employee.
What "computer use" means technically
"Computer use" is the industry term for AI systems that operate software the way humans do: perceiving the screen or page structure and taking actions like clicking, typing, and navigating. These systems combine vision/grounding models (locating buttons and fields), action models (planning the sequence of steps), and tool-use layers that fall back to APIs when more reliable than simulated clicks.
The browser is the natural beachhead because nearly all knowledge work already happens there. An agent that can drive a browser can, in principle, do any desk job a human does through a browser — which is why every major AI lab is racing to own this layer.
The browser as the AI battleground
This release sharpens a fight building for two years. The question is no longer "which chatbot is smartest" but "which agent actually does things" — and the browser is where it's being decided. At stake: browser vendors control the platform, AI labs build the agents that drive it, and millions of websites whose business models assume human visitors — human attention spans, click patterns, and susceptibility to advertising — face intermediation by software.
When agents browse on behalf of users, several things change at once: comparison shopping becomes frictionless (bad for sites relying on confusion, good for consumers), form-filling and checkout become automatable (good for conversion, alarming for fraud teams), and the website as a designed experience matters less than the website as structured data (good for whoever controls the agent layer, since they intermediate the relationship).
The security and permission questions
With great agency comes a list of questions founders should be asking — as builders on these capabilities and as operators whose products agents will visit:
- Permission granularity: What exactly can the agent touch? Reading tabs is one thing; sending messages and filling forms is another. Per-site, per-action, per-session permission models will determine how much users trust Work mode with real tasks.
- Credential exposure: An agent that fills forms and logs in handles passwords, payment details, and session tokens. How those are stored, scoped, and revoked is a first-order security design problem.
- Prompt injection via the web: A page the agent reads can contain hidden instructions. Agents that both read untrusted web content and take consequential actions are the prime target for this attack class; defenses — instruction hierarchies, action confirmation, domain allowlists — are still maturing.
- Accountability: When an agent books the wrong flight or sends the wrong message, who is liable — the user, OpenAI, or the website? Terms of service are racing to catch up.
- Bot detection vs. agent traffic: Websites must distinguish legitimate user-delegated agents from malicious bots. Expect access tiers: sites that welcome agents, sites that charge them, and sites that block them.
What founders should consider
First, distribution is consolidating around the agent layer. If ChatGPT becomes the default way people interact with the web, the agent — not your website's UX — becomes your product's front door. Design "agent-readable" experiences now: clean structured data, stable interfaces, and machine-readable policies for what agents may do.
Second, the moat question flips. When any agent can fill any form and compare any sites, defensibility moves away from workflow lock-in toward proprietary data, network effects, and trust.
Third, build for the permission economy. Users will pay for agents they trust with consequential actions. Transparency about what the agent did, easy undo, and clear audit trails are the differentiators separating agents people actually delegate to from demos people admire once.
Why it matters for founders and investors
Browser agents are the next platform shift, and platform shifts mint new winners. For investors, the categories to watch: agent infrastructure, agent security (permissions, audit, injection defense), and vertical agents owning specific workflows end-to-end. The ChatGPT Chrome release validates the category while threatening every startup whose plan was "we'll be the agent layer" — differentiation must come from vertical depth or from owning the trust layer.
For founders, assume agents as users: instrument your product for agent traffic, publish machine-readable terms for agent access, and design onboarding that works when the "user" is software. The companies that treat agents as first-class customers will capture the delegation economy; the ones that block them as bots will find themselves routed around.
Angeled.com connects founders with investors.
Sources: digit.in (Oct 16, 2026)
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



