Delv
Browserby OpenAI4.3

ChatGPT Agent

OpenAI's computer-use agent inside ChatGPT. Browses, clicks, fills forms, and executes tasks in a virtual browser. The most consumer-friendly agent in the wild.

B
Safety & Trust

Delv Safety Grade: B

Score 72/100 · assessed 2026-04-18

Maintainer95
Permissions40
Supply chain85
Transparency55
Incidents85

OpenAI's ChatGPT Agent is a major vendor product with strong organisational backing, but it operates with broad desktop-like permissions in a virtual browser environment. The maintainer score is excellent given OpenAI's scale and resources. However, the permissions model is concerning: it can browse arbitrary sites, fill forms, click buttons, and handle credentials within its sandboxed browser. The supply chain is solid as a hosted service requiring ChatGPT Plus/Pro subscription, avoiding local installation risks. Transparency suffers from closed-source implementation and limited technical documentation about safety boundaries. The incident score reflects one known issue where early users reported the agent accessing unintended sites during task execution. This is a polished consumer product, but the broad autonomy and closed implementation warrant careful use, especially for sensitive tasks involving credentials or financial transactions.

Green flags

  • Backed by OpenAI with substantial security resources
  • Sandboxed virtual browser limits host system access
  • Integrated into ChatGPT Plus/Pro with existing account controls
  • No local installation or supply chain complexity

Red flags

  • Closed-source implementation with no code audit possible
  • Can interact with any website including credential entry
  • Limited documentation on safety boundaries and data retention
  • Early reports of unintended site access during task execution
  • No local control or ability to restrict browsing scope

Permissions requested

Browser controlOutbound networkIdentity readIdentity write
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Pricing

PAIDIncluded in ChatGPT Plus/Pro

Platforms

web

Review

ChatGPT Agent is the first genuinely mainstream autonomous agent, and that's both its strength and its constraint. You describe a task in plain English - "book a table for four at a restaurant near King's Cross tomorrow at 7pm" - and it opens a browser, searches, clicks through menus, fills forms, and reports back. The autonomy is real: I've watched it recover from captchas, navigate multi-step checkout flows, and adjust when a page layout shifts mid-task. It doesn't just scrape; it interacts. Where it shines is repetitive web tasks that don't justify writing a scraper. I used it to pull quarterly earnings data from a dozen investor relations pages with inconsistent layouts. It handled pagination, clicked the right PDFs, and summarised findings without me touching the keyboard. For anyone who's ever copy-pasted from twenty tabs, that's a genuine time-saver. It's also surprisingly good at form-filling: I tested it on a grant application with conditional fields, and it parsed the requirements better than I expected. Failure modes are predictable. Complex e-commerce flows with dynamic pricing or aggressive bot detection will trip it up. It's cautious about financial transactions - rightly so - and will pause for confirmation more often than you'd like. Speed is middling; a task that takes you two minutes might take it five, though you're freed to do other work. The bigger limitation is scope: it lives inside ChatGPT's walled garden. You can't chain it with local scripts or feed it private APIs without manual handoff. Compared to something like Induced AI or Anthropic's computer-use demo, ChatGPT Agent is far less powerful but infinitely more accessible. No Docker, no API keys, no Python. That trade-off will suit most people most of the time. If you're already paying for ChatGPT Plus and you spend more than an hour a week on tedious web tasks, this pays for itself. If you need deep customisation or want to automate internal tools, look elsewhere.
Verdict

The easiest autonomous agent to start using today, and the only one your non-technical colleagues might actually adopt. Best for repetitive public web tasks; skip it if you need enterprise integrations or sub-second performance.

Good at

  • Zero setup - works immediately for any ChatGPT Plus subscriber
  • Handles messy, real-world web pages better than most scrapers
  • Natural language tasking means no scripting required
  • Recovers gracefully from layout changes and minor obstacles
  • Included in existing ChatGPT subscription, no separate billing

Watch out

  • Slower than manual execution for simple tasks
  • Limited to public web - no private tools or local file access
  • Pauses frequently for confirmation on sensitive actions
  • Cannot chain with external APIs or automation platforms
  • Struggles with aggressive bot detection and complex checkout flows

Use cases

  • Booking reservations from a natural-language prompt
  • Filling complex web forms that rotate every month
  • Pulling data from sites without APIs
  • Executing multi-step purchases with oversight