Database/AI/ML frameworks & serving

SGLang: LoRA tensor-load endpoint skips admin auth, letting anyone on the network alter the served model
Impact
The POST /load_lora_adapter_from_tensors endpoint is missing the admin authorization check that its sibling admin endpoints have, so it accepts requests without an API key even when one is configured. An attacker who can reach the server's HTTP port can load a LoRA adapter built from tensors they supply and change what the served model returns to every client. The record says that chained with a separately documented deserialization bypass, this becomes unauthenticated remote code execution as root on the serving host, which is a GPU node holding model weights and possibly other tenants' traffic.
Who can reach it
Anyone who can reach the SGLang HTTP API port. No authentication needed, and setting an API key does not protect this endpoint.
What to do
Upgrade SGLang to a build that contains commit 7b5fb39 (the record lists affected versions through 0.5.23 and does not name a fixed release), then restart each serving process. Until then, keep the SGLang port off untrusted networks and block /load_lora_adapter_from_tensors at the reverse proxy or gateway in front of it.
References
Related entries
- vLLM: out-of-range prompt token ids crash the shared EngineCore when tokenizer init is skippedCVE-2026-130685 · vLLM OpenAI-compatible server (--skip-tokenizer-init token id validation)High
- vLLM: no per-request media budget lets one request with many image URLs exhaust the frontendCVE-2026-130686 · vLLM OpenAI-compatible server (multimodal remote image fetching)High
- KubeAI (Ollama engine controller): Injection in `ollamaStartupProbeScript()`CVE-2026-34940 · KubeAI (Ollama engine controller)High
- Flowise: public chatflow endpoints return raw flowData including plaintext API keysCVE-2026-41278 · Flowise (GET /api/v1/public-chatflows/:id and public-chatbotConfig response sanitization)High
- Nocturne Memory: empty API_TOKEN disables auth, exposing the agent knowledge graph to the LANCVE-2026-44830 · Nocturne Memory (BearerTokenAuthMiddleware when API_TOKEN is unset)High
- vLLM: user-supplied structured-output regex compiles with no timeout, hanging an inference workerCVE-2026-55574 · vLLM structured_outputs.regex (xgrammar / outlines grammar compilation)High
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.