RealFeedApp SGLang fixed KV cache exposure in GLM-5.3
A vulnerability was found allowing access to KV cache from one request by another when using SGLang PD. Developers implemented a deferred memory release mechanism and checksum verification to fix the issue.
- Field
- computer security
- What they did
- Researchers discovered a vulnerability in the SGLang system where cache data (KV cache) from one request could incorrectly leak into the cache of the next request, violating data isolation. This happened because the system did not always correctly determine when a cache slot is considered free for rewriting.
- Why it matters
- Fixing this vulnerability is necessary to protect user privacy and data security in distributed systems where each client must have access only to their own data.
#sglang#kv cache#security#llm#inference
