SGLang through 0.5.20 in prefill/decode disaggregation mode fails to validate duplicate bootstrap_room fields in /generate requests with Mooncake KV transfer backend. Unauthenticated attackers can send concurrent requests with identical bootstrap_room values to crash scheduler processes or hang other users requests until transfer timeout.
Weakness
The product uses multiple resources that can have the same identifier, in a context in which unique identifiers are required.
Potential Mitigations
References
- https://github.com/sgl-project/sglang
- https://github.com/sgl-project/sglang/blob/v0.5.20/python/sglang/srt/disaggregation/common/conn.py#L392-L393
- https://github.com/sgl-project/sglang/blob/v0.5.20/python/sglang/srt/disaggregation/decode.py#L942
- https://github.com/sgl-project/sglang/blob/v0.5.20/python/sglang/srt/disaggregation/mooncake/conn.py#L2760-L2764
- https://github.com/sgl-project/sglang/blob/v0.5.20/python/sglang/srt/managers/io_struct.py#L281
- https://github.com/sgl-project/sglang/issues/40125
- https://www.vulncheck.com/advisories/sglang-through-0.5.20-denial-of-service-via-duplicate-bootstrap-room
- https://github.com/sgl-project/sglang/issues/40125