You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Uses OpenAI image generation models such as DALL·E, `gpt-image-1`, and `gpt-image-1.5` to generate images from text prompts.
15
+
Uses OpenAI image generation model `gpt-image` to generate images from text prompts.
18
16
19
17
Up to the winter 2026 OpenAI models are one of the most powerful image generation models available, especially GPT-Image-1.5, which is why we started with them.
To use input images, the adapter downloads every URL passed in `inputFiles` from your server. If those URLs can be influenced by a user (for example they are built from a record field which the user can edit, or come from `attachFiles` of a plugin), an attacker can point them at internal addresses like `http://169.254.169.254/latest/meta-data/` or `http://localhost:9200/` and use your backend as a proxy to your private network — a classic SSRF.
50
+
51
+
`attachImagesAllowedHosts` closes this hole: only listed hosts are downloaded, anything else throws before the request is made. Always set it to the storage you actually serve images from:
Matching is case-insensitive: `example.com` matches that host only, while `.example.com` (or `*.example.com`) matches it and all its subdomains.
28
62
29
-
This model is the top of the Nano Banana line as of 2026, combining the speed of the Flash series with the improved detail of version 3.1. The adapter lets you integrate the advanced capabilities of previous models into your interface and generate precise visuals even for specific or complex prompts.
63
+
> ⚠️ When `attachImagesAllowedHosts` is not set, images are downloaded from **any** host. Leave it unset only if the adapter never receives URLs which depend on user input.
0 commit comments