The “error browsing” may be further “untrusted” actions or untrusted http document types. You’ll need a verified builder domain, schema, authentication, privacy policy, etc. and the ability to then publish to “everyone” to then make those custom actions.
Then the tool “browsing” is going to scrape page data.
I still shouldn’t expect that ChatGPT vision can “see” the contents of an image; The only way that an image can be perceived is to have it be part of a user message to the GPT-4 model that supports computer vision. There is no other path of loading or analyzing images except on demand of the user (or an API user placing those images into a user role message).