diff --git a/docs/changelog/154972.yaml b/docs/changelog/154972.yaml
new file mode 100644
index 0000000000000..bccf36f0ffb5e
--- /dev/null
+++ b/docs/changelog/154972.yaml
@@ -0,0 +1,6 @@
+area: ES|QL
+issues:
+ - 154866
+pr: 154972
+summary: Addressing inference timeout issues on inference operator tests
+type: bug
diff --git a/x-pack/plugin/esql/src/main/java/org/elasticsearch/xpack/esql/inference/InferenceOperator.java b/x-pack/plugin/esql/src/main/java/org/elasticsearch/xpack/esql/inference/InferenceOperator.java
index a6babe277a6e8..320fa15889be9 100644
--- a/x-pack/plugin/esql/src/main/java/org/elasticsearch/xpack/esql/inference/InferenceOperator.java
+++ b/x-pack/plugin/esql/src/main/java/org/elasticsearch/xpack/esql/inference/InferenceOperator.java
@@ -377,7 +377,10 @@ public BulkInferenceRequestItem pollNextRequest() {
             }
 
             synchronized (checkpoint) {
-                if (requestItemIterator.hasNext() == false) {
+                // Re-check under the lock: completeIfFinished() releases the request iterator (and its backing block) while holding
+                // this lock when a failure occurs. Without re-checking, a concurrent poller that passed the guard above could invoke
+                // requestItemIterator.next() on an already-released block. See https://github.com/elastic/elasticsearch/issues/154866
+                if (hasFailure() || completed.get() || requestItemIterator.hasNext() == false) {
                     return null;
                 }