diff --git a/prover/circuits/pi-interconnection/keccak/prover/maths/common/vector/vector.go b/prover/circuits/pi-interconnection/keccak/prover/maths/common/vector/vector.go
index 4e96d19f511..458b3b9aef8 100644
--- a/prover/circuits/pi-interconnection/keccak/prover/maths/common/vector/vector.go
+++ b/prover/circuits/pi-interconnection/keccak/prover/maths/common/vector/vector.go
@@ -36,6 +36,15 @@ func ScalarMul(res, vec []field.Element, scalar field.Element) {
 func ScalarProd(a, b []field.Element) field.Element {
 	// The length checks is done by gnark-crypto already
 	a_ := fr.Vector(a)
+	// Workaround for gnark-crypto AVX-512 innerProdVec overread:
+	// The assembly reads 4 bytes past the last element of the receiver via
+	// VPMULUDQ.BCST at offset 28 (8-byte load from byte 28 of a 32-byte element).
+	// When the allocation is page-aligned this causes SIGSEGV. Ensure extra capacity.
+	if len(a) > 0 && cap(a) == len(a) {
+		padded := make(fr.Vector, len(a)+1)
+		copy(padded, a_)
+		a_ = padded[:len(a)]
+	}
 	res := a_.InnerProduct(fr.Vector(b))
 	return res
 }
diff --git a/prover/circuits/pi-interconnection/keccak/prover/maths/field/gen.go b/prover/circuits/pi-interconnection/keccak/prover/maths/field/gen.go
index ac4983dce1f..a1a1294a589 100644
--- a/prover/circuits/pi-interconnection/keccak/prover/maths/field/gen.go
+++ b/prover/circuits/pi-interconnection/keccak/prover/maths/field/gen.go
@@ -16,7 +16,9 @@ func ParBatchInvert(a []Element, numCPU int) []Element {
 		numCPU = runtime.NumCPU()
 	}
 
-	res := make([]Element, len(a))
+	// Allocate with +1 capacity to prevent gnark-crypto AVX-512 innerProdVec
+	// overread when the result is subsequently used in InnerProduct.
+	res := make([]Element, len(a), len(a)+1)
 
 	parallel.Execute(len(a), func(start, stop int) {
 		subRes := BatchInvert(a[start:stop])