Database/AI/ML frameworks & serving
vLLM: model config.json can redirect the served model and tokenizer to another local checkpoint
Impact
A malicious model publisher can set speculators_config verifier.name_or_path in a model's config.json to point at a different local checkpoint directory (path traversal). When an operator serves that model, vLLM loads the other checkpoint's tokenizer and model instead. The attacker can then read that checkpoint's tokenizer vocabulary through unauthenticated /detokenize requests and send queries to a private model they should not reach. On shared inference nodes that keep several customers' or teams' checkpoints on local disk, one untrusted model can expose another tenant's model.
Who can reach it
Someone who publishes a model that an operator then loads into vLLM (user interaction: the model has to be deployed). After that, any client that can reach the vLLM HTTP endpoint can use /detokenize without authentication.
What to do
Upgrade vLLM to 0.33.0 or later and restart the serving processes or pods. Until then, only serve models from trusted publishers, review config.json speculators_config fields before deployment, and keep other checkpoints out of paths the vLLM process can read.
References
Related entries
- BentoML OpenLLM 0.6.30 (async_run_command in src/openllm/common.py): A model repository directory name flows unescapedCVE-2026-15035 · BentoML OpenLLM 0.6.30 (async_run_command in src/openllm/common.py)Medium
- RAGFlow: unvalidated base_url in set_api_key lets authenticated users make server-side requestsCVE-2026-49951 · RAGFlow (POST /v1/llm/set_api_key base_url handling)Medium
- JupyterHub: unauthenticated logins write unbounded usernames to the log, exhausting storageCVE-2026-54338 · JupyterHub form-based login authenticators (failed-login logging)Medium
- vLLM: malformed JSON to the OpenAI-compatible endpoints returns server paths and versionsCVE-2026-73555 · vLLM OpenAI-compatible API server (validation_exception_handler, sanitize_message)Medium
- vLLM: attacker-supplied structured-output regex pins a CPU core and stalls the engine pathCVE-2026-73556 · vLLM structured outputs, lm-format-enforcer backend (structured_outputs.regex)Medium
- vLLM: integer overflow in the activation CUDA kernel leaks another batched request's outputCVE-2026-73558 · vLLM CUDA activation kernels (act_and_mul_kernel, activation_kernels.cu)Medium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.