NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · vLLM

CVE-2026-73555: vLLM: Unauthenticated Internal Path and Username Disclosure via Validation Error Messages

mediumCVSS 3.1 5.3GHSA-hwrm-c4cx-rf4j

vLLM's OpenAI-compatible server leaks internal file paths in its error responses. A single malformed JSON request makes the validation error handler return text with the install path, OS username and line numbers. No authentication is needed. Versions below 0.26.0 are affected. Fixed in 0.26.0.

Published Sep 4, 2026 · Analysis Nov 1, 2026

What an attacker can do

When a request body is invalid, FastAPI raises a Pydantic RequestValidationError. The validation_exception_handler in vllm/entrypoints/openai/server_utils.py turned it into a string with str(exc). That string includes the file path and line number of the endpoint handler. The existing sanitize_message() removed memory addresses such as 0x7f... but not File "...", line N patterns, so the text reached the client.

One bad request to any POST endpoint that takes JSON, such as /v1/chat/completions, /v1/completions, /tokenize or /detokenize, shows:

  • the OS user running vLLM and its home directory, for example /home/ubuntu/;
  • the virtualenv path and Python version;
  • internal module paths, handler names and line numbers.

The handler names and line numbers identify the exact vLLM build, even if the /version endpoint is disabled. This is reconnaissance only: confidentiality impact is low (5.3), with no code execution or data access. It helps an attacker pick follow-on exploits for an exposed inference server.

Who is affected

vLLM below 0.26.0 with the OpenAI-compatible API server reachable by the attacker. The reporter tested 0.20.1 on Python 3.12, and said the bug is independent of the model. Check your version:

pip show vllm

To test your own server, send an invalid JSON body to /v1/chat/completions and look for File and line text in the error message.

vllm

Ecosystem
PyPI
Affected
< 0.26.0
Fixed in
0.26.0
Check
pip show vllm

How to fix it

Upgrade to vLLM 0.26.0.

pip install --upgrade vllm==0.26.0

Pull request 46415 builds the validation error message from exc.errors() instead of str(exc), so no path enters the response. It also makes sanitize_message strip traceback-style frame lines and absolute filesystem paths. As a workaround, the advisory suggests a reverse proxy that rewrites error bodies to remove file paths, and calls it fragile.

How to tell if you were hit

Review access logs for 400 responses on /v1/chat/completions, /v1/completions, /tokenize and /detokenize from unknown clients, especially bursts of malformed bodies across several endpoints, which suggests fingerprinting. Capture your server's response to a deliberately bad request: if it contains File " or a home directory path, the instance is unpatched.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

When the vLLM API receives a malformed request (e.g., invalid JSON or missing required fields), FastAPI raises a Pydantic RequestValidationError. The validation_exception_handler in vllm/entrypoints/openai/server_utils.py converts this exception to a string via str(exc), which includes the internal file path and line number of the handler function. The existing sanitize_message() function in vllm/entrypoints/utils.py strips memory addresses (e.g., 0x7f...) but does not strip File "...", line X patterns. The result is a user-facing HTTP response that leaks internal system information.

Impact

An unauthenticated attacker can extract the following with a single malformed request:

  • OS username running the vLLM process (e.g., ubuntu)
  • Home directory path (e.g., /home/ubuntu/)
  • Virtual environment path (e.g., vllm-env/)
  • Python version (e.g., 3.12)
  • Internal package structure and line numbers (e.g., vllm/entrypoints/openai/chat_completion/api_router.py)
  • Handler function names per endpoint, enabling precise version fingerprinting

This information aids attackers in constructing targeted exploits: environment paths narrow the attack surface, and handler function names + line numbers enable exact version identification even when the /version endpoint is disabled.

All POST endpoints that accept JSON bodies are affected, including /v1/chat/completions, /v1/completions, /tokenize, and /detokenize.

Workarounds

Deploying vLLM behind a reverse proxy that rewrites error response bodies to strip file paths would mitigate this, though it is fragile.

Remediation Recommendation

Two possible fixes (either suffices):

Option A — Fix `validation_exception_handler`: Construct the error message from exc.errors() (the structured Pydantic error list) rather than str(exc). This avoids the traceback-style string entirely.

Option B — Fix `sanitize_message`: Add a regex to strip File "...", line \d+ patterns, similar to how memory addresses are already stripped:

import re
msg = re.sub(r'File ".*?", line \d+, in \w+', '[internal]', msg)

Option A is preferred as it addresses the root cause rather than filtering symptoms.

Environment Tested

  • vLLM 0.20.1 (pip install, latest stable as of May 2026)
  • Python 3.12
  • Ubuntu 22.04
  • Model: Qwen/Qwen2-0.5B (text-only; bug is model-independent)

This was fixed here: https://github.com/vllm-project/vllm/commit/e87521626f

Credited to biecho, jperezdealgaba.

Questions

Does CVE-2026-73555 need authentication?

No. The advisory describes an unauthenticated attacker sending one malformed request to a POST endpoint that accepts JSON.

What can an attacker learn?

The OS username, home and virtualenv paths, Python version, internal module paths, handler names and line numbers. The last two let them identify the exact vLLM version.

Does turning off the /version endpoint help?

No. The advisory says handler names and line numbers in the error text still reveal the version.

Is there a workaround?

A reverse proxy that strips file paths from error bodies would help, but the advisory calls that fragile. Upgrade to 0.26.0.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More vLLM vulnerabilities

All vLLM entries