# Agentic Browser

> An agentic browser is a web browser or extension with a built-in AI assistant that can read the pages you have open and act in them: click, type and fill forms.

An agentic browser is a web browser or extension whose built-in AI assistant can read the open pages and act in them: click, type, fill forms and work across tabs. Current examples are Perplexity Comet, Dia and Claude in Chrome; ChatGPT Atlas was retired in August 2026. The core risk is prompt injection from page content.

**An agentic browser is a web browser, or a browser extension, whose built-in AI assistant can both read the pages you have open and act inside them: click, type, fill forms, and carry a task across several tabs.** It is [computer use](/glossary/computer-use) narrowed to the browser, where the agent already has your logged-in sessions.

The category has two shapes. **Standalone browsers** ship the assistant as part of a Chromium build. Perplexity's [Comet](/tools/perplexity-comet) is "a Chromium-based browser that integrates advanced AI capabilities powered by Perplexity," free on Mac, Windows, iOS and Android, with a Comet Assistant that acts in tabs and shows you each action it takes. [Dia](/tools/dia-browser), now part of Atlassian, is built around a chat sidebar that reads your tabs and connected work apps, on macOS and, since October 6, 2026, Windows; as of October 2026 Dia advertises no autonomous click-and-type mode. OpenAI's ChatGPT Atlas, a macOS browser launched in October 2025 with an agent mode, was retired on August 9, 2026. **Extensions** add the agent to a browser you already use. Anthropic's [Claude in Chrome](/tools/claude-for-chrome) "reads the page you're signed in to, then clicks, types, and fills forms," and became generally available on every paid Claude plan on August 26, 2026.

The defining risk is [prompt injection](/glossary/prompt-injection). Because the agent reads page content as input, instructions planted in that content can hijack it. Brave's researchers showed in August 2025 that hidden text in a Reddit comment could make Comet read the user's email and a one-time passcode, and in October 2025 that "malicious instructions embedded as nearly-invisible text within the image are processed as commands." Vendors have responded with layered defenses: Perplexity trained classifiers to detect malicious instructions, Anthropic validates each Claude in Chrome action with a safety classifier, and OpenAI's own assessment was that prompt injection "is unlikely to ever be fully 'solved'." The [best AI browsers in 2026](/guides/comparisons/best-ai-browsers-2026) roundup compares the current options.

---

_Source: https://agentscamp.com/glossary/agentic-browser — Term on AgentsCamp._
