The with context constructors of the stream encoders and decoders used the borrowed context as it was. When a stream was abandoned on a context, for example an encoder dropped before finish() after an error, whatever it left there was carried into the next stream: the input buffered by the previous encoder (up to 128 KiB, or all of it with zstdmt) became the start of the next frame, and a decoder returned undelivered output of the previous stream.

An application reusing one context for data that must stay separate, such as different users' requests, could disclose one user's plaintext to another. Whoever supplies the data can cause the abandonment with invalid input.

The flaw was corrected by resetting the context's session in with context.