Cross-tenant DoS via uninitialized LogprobsTensors with prompt_logprobs
GHSA-rhcx-5729-88vg · published 2026-10-09
- Severity
- High
- CVSS score
- 7.7
- Chance of exploitation (EPSS)
- No CVE
- Fixed in
- 0.31.0
Remote-trigger crash of EngineCore on a single malformed request. Scope change: the crash kills every in-flight request on the engine, not just the offender's. Co-tenant traffic and in-progress batches are dropped. Trivial to trigger: warm a prefix (or rely on natural cache reuse), then send a request with prompt_logprobs set. Repeatable as a kill-loop. Highest real-world exposure: lm-eval-harness multiple-choice workloads (MMLU, ARC, HellaSwag, BBH, PIQA) against vLLM + LMCache. MC tasks require prompt_logprobs by definition and share per-subject prefixes by construction; item 2 of any subject is a guaranteed crash on a connector-equipped server.
Upgrade vllm to 0.31.0 or later. If you start it from an MCP config with npx or uvx, pin that version so an old cached copy is not used.
Replace torch.empty() with torch.zeros() for token IDs and torch.full(-inf) for logprobs in LogprobsTensors.empty_cpu(). Zero is a valid token ID that decodes safely; -inf is the correct "not computed" sentinel and matches the existing convention for out-of-top-k positions. All non-connector code paths overwrite every position, so the initial values are dead writes in the common case — no behavior change for any previously-working code path. The upstream commit should carry a Fixes: trailer pointing at the regression commit, to let maintainers triage backports automatically:
Until the patch is deployed, any one of the following mitigates the issue: Disable KV connectors on the affected engine (drop --kv-transfer-config and any kv_transfer_config setting). Local prefix caching is unaffected and remains available. Strip prompt_logprobs, logprobs, and echo from inbound requests at an API gateway or middleware in front of vLLM. These are the only request-level switches that reach the crash path. Apply the one-line patch above (torch.empty → torch.zeros / torch.full(-inf) in vllm/v1/outputs.py). No new dependencies, no behavior change on any non-connector path.
| Package | Registry | Affected | Fixed in |
|---|---|---|---|
| vllm | PyPI | >= v0.8.5, < 0.31.0 | 0.31.0 |
- HightsParticles: Unbounded HTTP Request Body Buffering in @tsparticles/mcp-server Leads to Remote DoS@tsparticles/mcp-server
- HighMCP TypeScript SDK: HTTP server transports and handlers did not limit request body size or JSON-RPC batch length@modelcontextprotocol/hono, @modelcontextprotocol/node, @modelcontextprotocol/sdk, @modelcontextprotocol/server
- HighMCP TypeScript SDK: One HTTP request within the default size limit could exhaust server memory during validation@modelcontextprotocol/sdk, @modelcontextprotocol/server
- HighHTTP server transports and OAuth endpoints read request bodies with no size limitmcp, mcp
- HighLMDeploy vulnerable to arbitrary code execution via eval() of untrusted quant_dtype in model config loadinglmdeploy
- HighRMCP: Unauthenticated permanent session-table leak in rmcp Streamable HTTP server transport leads to remote denial-of-servicermcp
- CriticalvLLM: OpenAI auth bypassvllm
- CriticalvLLM has RCE In Video Processingvllm
- CriticalvLLM Allows Remote Code Execution via PyNcclPipe Communication Servicevllm
- CriticalvLLM Vulnerable to Remote Code Execution via Mooncake Integrationvllm
- CriticalCVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0vllm
- CriticalvLLM Deserialization of Untrusted Data vulnerabilityvllm
The impact, patch and workaround text comes from the advisory, shortened (published by vLLM). Proof-of-concept and reproduction sections and code are left out; the original advisory has them.