Seen in several production clusters running 8.3 (thanks to the changes in #87881) but this also explains #86059 in earlier versions:

[instance-0000000011] negative value in circuit breaker: [eql_sequence,limit=8589934592/8gb,estimated=-7961288

My guess is that somehow this circuit breaker is releasing bytes it didn't already hold. Seems pretty rare, and not covered by tests since it would trip assertions there.