diff --git a/double-conversion/fast-dtoa.cc b/double-conversion/fast-dtoa.cc
index d7a23984..86fffd15 100644
--- a/double-conversion/fast-dtoa.cc
+++ b/double-conversion/fast-dtoa.cc
@@ -452,6 +452,15 @@ static bool DigitGenCounted(DiyFp w,
   *kappa = divisor_exponent_plus_one;
   *length = 0;
 
+  // With requested_digits <= 0 no digit can be produced. The integral loop
+  // below decrements requested_digits and only stops on an exact 0, so a
+  // non-positive value never trips that break and the loop emits every digit
+  // of 'integrals' past the end of the buffer. DoubleToAscii guards PRECISION
+  // with 0 digits before reaching here, but a direct FastDtoa caller does not.
+  if (requested_digits <= 0) {
+    return false;
+  }
+
   // Loop invariant: buffer = w / 10^kappa  (integer division)
   // The invariant holds for the first iteration: kappa has been initialized
   // with the divisor exponent + 1. And the divisor is the biggest power of ten
diff --git a/test/cctest/test-fast-dtoa.cc b/test/cctest/test-fast-dtoa.cc
index 5fb8975b..b059d412 100644
--- a/test/cctest/test-fast-dtoa.cc
+++ b/test/cctest/test-fast-dtoa.cc
@@ -296,6 +296,20 @@ TEST(FastDtoaPrecisionVariousDoubles) {
   CHECK(status);
   CHECK_EQ("7989", buffer.start());
   CHECK_EQ(192, point);
+
+  // requested_digits <= 0 must not write any digit. Fence the one-byte buffer
+  // with a sentinel and confirm FastDtoa fails without touching past it.
+  char tiny_container[2];
+  tiny_container[1] = 0x7f;
+  Vector<char> tiny(tiny_container, 1);
+  status = FastDtoa(1234567.0, FAST_DTOA_PRECISION, 0, tiny, &length, &point);
+  CHECK(!status);
+  CHECK_EQ(0, length);
+  CHECK_EQ(0x7f, static_cast<int>(tiny_container[1]));
+  status = FastDtoa(1234567.0, FAST_DTOA_PRECISION, -3, tiny, &length, &point);
+  CHECK(!status);
+  CHECK_EQ(0, length);
+  CHECK_EQ(0x7f, static_cast<int>(tiny_container[1]));
 }