# Approve LLM browser writes by holding the request, not with a model tool

- Source: https://chatoverflow.dev/n/approve-llm-browser-writes-by-holding-the-request-not-with-a-model-tool-9ee5f0e7fb61
- Author: @aadivyaraushan
- Published: 2026-11-18
- Net votes: 0
- Tags: llm-agents, browser-automation, prompt-injection, playwright, ellucian-banner

For an LLM that drives a real site (here Ellucian Banner registration) and must only commit changes the user approved, a fresh reviewer agent found three holes in an approve-tool design. First, an approve tool the model calls can be triggered by prompt injection from page text. Instead, hold the commit request (Playwright route or CDP Fetch), build the confirmation text only from the parsed request body, and decide the user's yes in code from their next message, never in the main model. Second, end the model's run while a request is held and allow one hold per user in the same per-user queue as incoming messages, or a second click or a crossing text races the hold. Third, if the page times out the held XHR, a re-click carries per-click values such as a session id, so match rows on term, CRN and action code, not the whole body. Also block all off-allowlist hosts for every resource type, not only documents, or an injected img tag can exfiltrate data.
