diff --git a/cuda_bindings/cuda/bindings/_internal/strdecode.py b/cuda_bindings/cuda/bindings/_internal/strdecode.py
new file mode 100644
index 00000000000..9c723fe9d97
--- /dev/null
+++ b/cuda_bindings/cuda/bindings/_internal/strdecode.py
@@ -0,0 +1,39 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+"""Decode C strings returned by CUDA libraries with actionable failure context."""
+
+# Cap sized for the #2118 mojibake without flooding exception text.
+_PREVIEW_MAX_BYTES = 64
+
+
+def _bounded_hex_preview(data: bytes, max_bytes: int = _PREVIEW_MAX_BYTES) -> str:
+    # Bytes after the first NUL are not part of the returned C string. The
+    # marker is explicit so truncation cannot be misread as an empty value.
+    nul = data.find(b"\x00")
+    nul_stopped = nul != -1
+    visible_end = len(data) if not nul_stopped else nul
+    snippet_end = min(visible_end, max_bytes)
+    snippet = data[:snippet_end]
+    body = snippet.hex(" ") if snippet else ""
+    parts = []
+    if snippet_end < visible_end:
+        parts.append(f"+{visible_end - snippet_end} more")
+    if nul_stopped:
+        parts.append(f"stopped at NUL@{nul}")
+    suffix = f" ...({'; '.join(parts)})" if parts else ""
+    return f"<{visible_end} bytes; hex='{body}'{suffix}>"
+
+
+def decode_c_str(data: bytes, api_name: str) -> str:
+    """Decode ``data`` as UTF-8, or raise ``UnicodeDecodeError`` with ``api_name`` and a bounded hex preview in ``reason``.
+
+    Internal API. ``api_name`` is trusted caller input and embedded verbatim.
+    """
+    try:
+        return data.decode("utf-8")
+    except UnicodeDecodeError as e:
+        # Same exception type, not a subclass, so existing handlers still catch.
+        preview = _bounded_hex_preview(data)
+        reason = f"{e.reason} (returned by {api_name}; bytes={preview})"
+        raise UnicodeDecodeError(e.encoding, e.object, e.start, e.end, reason) from e
diff --git a/cuda_bindings/tests/test_strdecode.py b/cuda_bindings/tests/test_strdecode.py
new file mode 100644
index 00000000000..2d68f6cf4c1
--- /dev/null
+++ b/cuda_bindings/tests/test_strdecode.py
@@ -0,0 +1,80 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+import pytest
+
+from cuda.bindings._internal.strdecode import _bounded_hex_preview, decode_c_str
+
+WSL_MOJIBAKE_PREFIX = b"\xf8\x9a\x80\x80\xaf"
+
+
+def test_valid_utf8_passthrough():
+    assert decode_c_str(b"hello world", "fakeApi") == "hello world"
+
+
+def test_invalid_bytes_raise_unicode_decode_error():
+    with pytest.raises(UnicodeDecodeError):
+        decode_c_str(WSL_MOJIBAKE_PREFIX, "nvmlSystemGetProcessName")
+
+
+def test_failure_reason_includes_api_name():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(WSL_MOJIBAKE_PREFIX, "nvmlSystemGetProcessName")
+    assert "nvmlSystemGetProcessName" in excinfo.value.reason
+
+
+def test_failure_reason_includes_hex_preview():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(WSL_MOJIBAKE_PREFIX, "nvmlSystemGetProcessName")
+    assert "f8 9a 80 80 af" in excinfo.value.reason
+
+
+def test_failure_chains_original_error():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(b"\xf8", "fakeApi")
+    assert isinstance(excinfo.value.__cause__, UnicodeDecodeError)
+
+
+def test_failure_preserves_codec_and_position():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(b"\xf8\x9a", "fakeApi")
+    assert excinfo.value.encoding == "utf-8"
+    assert excinfo.value.start == 0
+    assert excinfo.value.end == 1
+
+
+def test_preview_stops_at_first_nul():
+    preview = _bounded_hex_preview(b"\xf8\xf8\x00trailing junk")
+    assert "f8 f8" in preview
+    assert "trailing" not in preview
+    assert "<2 bytes;" in preview
+    assert "stopped at NUL@2" in preview
+
+
+def test_preview_caps_long_buffers():
+    preview = _bounded_hex_preview(b"\xf8" * 200, max_bytes=8)
+    assert "f8 f8 f8 f8 f8 f8 f8 f8" in preview
+    assert "+192 more" in preview
+    assert "stopped at NUL" not in preview
+
+
+def test_preview_combines_truncation_and_nul_markers():
+    preview = _bounded_hex_preview(b"\xf8" * 20 + b"\x00rest", max_bytes=8)
+    assert "+12 more" in preview
+    assert "stopped at NUL@20" in preview
+
+
+def test_failure_preview_stops_at_embedded_nul_even_with_bad_bytes_before():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(b"\xf8\x9a\x00ignored_after_nul", "fakeApi")
+    reason = excinfo.value.reason
+    assert "f8 9a" in reason
+    assert "ignored_after_nul" not in reason
+
+
+def test_failure_message_stays_bounded_for_long_garbage():
+    with pytest.raises(UnicodeDecodeError) as excinfo:
+        decode_c_str(b"\xf8" * 1024, "fakeApi")
+    reason = excinfo.value.reason
+    assert "+960 more" in reason
+    assert len(reason) < 500