Hi OpenAI Team,
Right now, web browsing in ChatGPT is strictly read-only. It can find the steps to renew a passport, pay a parking ticket, or change an address, but it can’t actually complete the process for you. When people try to do these mundane tasks themselves, they constantly fall for predatory SEO traps (sponsored ads, lookalike DMV portals, and third-party fee-skimming sites). If ChatGPT had pre-vetted task flows for authoritative domains, it could turn passive web search into active read/write execution, ensuring users navigate the legitimate portal while completely cutting out the scam ecosystem.
Current “Computer Use” paradigms attempt to solve part of this, but an agent should never be fed the sensitive data needed to renew a passport or buy a ticket if that data lives in a decrypted chat context or visual frame buffer. The blocker has always been the security risk of handling credentials and leaking them into the context window, but that is solvable by decoupling the AI’s reasoning from the actual data payload. The architecture doesn’t need to be a clunky, token-heavy vision model streaming screenshots of a mouse. Instead, OpenAI could run a dedicated, encrypted secondary plane that’s conceptually similar to a PCI-compliant tokenization vault or hardware security module (HSM) and that operates completely outside the LLM’s inference context window. The model acts strictly as the control plane: it identifies the form fields, handles the navigation logic, and issues an abstract instruction like “populate payment method.” The secondary plane then handles the actual decryption and injects the credentials directly into the target site’s checkout session over TLS. The raw credit card, CVV, or SSN never enters the transformer’s attention heads, never appears in prompt logs, and is structurally immune to prompt-injection extraction because the model itself literally cannot read the field, and is end-to-end AES-256 encrypted (if built properly).
From a user experience standpoint, this works seamlessly across both phone and desktop. Users could either save a card into an encrypted wallet profile within their account for one-click delegated tasks, or simply have the app prompt them on the fly with a native Apple Pay, Google Pay, or biometric sheet to authorize the charge. It turns ChatGPT into an authentic administrative copilot that can safely execute real-world transactions without turning user sessions into a security liability. And if Cloudflare or reCAPTCHA trips along the way? Just pass that through as a quick native popup too (we’re all happy to click the traffic lights for three seconds if ChatGPT is doing the other twenty minutes of paperwork).