| Titre | zhayujie chatgpt-on-wechat 2.1.0 Server-Side Request Forgery (SSRF) (CWE-918) |
|---|
| Description | # Technical Details
A Server-Side Request Forgery vulnerability exists in the `[Vision._download_to_data_url]` method in `agent/tools/vision/vision.py` of chatgpt-on-wechat.
The application fails to validate attacker-controlled remote image URLs before performing a server-side fetch. `Vision.execute()` accepts the `image` argument, passes it into `_build_image_content(image)`, and HTTP(S) inputs are routed directly into `_download_to_data_url()` which executes `requests.get(url, timeout=30)` without loopback, RFC1918, link-local, DNS-resolution, or redirect-chain validation.
# Vulnerable Code
File: `agent/tools/vision/vision.py`
Method: `Vision.execute`, `_build_image_content`, `_download_to_data_url`
Why: The code accepts arbitrary `http://` and `https://` image URLs and fetches them directly with `requests.get()` before any destination policy is enforced, allowing requests to internal or localhost resources.
# Reproduction
1. Configure at least one vision-capable provider so execution reaches the fetch path.
2. Run the supplied verification script with a canary URL such as `http://127.0.0.1:<port>/canary.png`.
3. Observe the local canary server receives a GET request from `python-requests` before the provider later fails upstream.
# Impact
- Allows attackers to induce server-side requests from the agent host to localhost or internal-network targets.
- Can disclose or interact with internal HTTP services reachable from the host running the agent. |
|---|
| La source | ⚠️ https://github.com/zhayujie/CowAgent/issues/2878 |
|---|
| Utilisateur | Eric-e (UID 97581) |
|---|
| Soumission | 11/06/2026 10:39 (il y a 1 mois) |
|---|
| Modérer | 13/07/2026 19:28 (1 month later) |
|---|
| Statut | Accepté |
|---|
| Entrée VulDB | 378130 [zhayujie chatgpt-on-wechat CowAgent jusqu’à 2.1.1 Vision Tool vision.py Vision._download_to_data_url image élévation de privilèges] |
|---|
| Points | 20 |
|---|