diff --git a/enroll/ignore.py b/enroll/ignore.py
index 563a9ee..7919d1b 100644
--- a/enroll/ignore.py
+++ b/enroll/ignore.py
@@ -180,8 +180,21 @@ HIGH_CONFIDENCE_SECRET_PATTERNS = [
     ),
     # HTTP(S) Authorization / Proxy-Authorization header values carrying a
     # bearer/basic/digest credential.
+    #
+    # This is high-confidence and therefore scanned against the RAW bytes so a
+    # *commented-out* populated header (a real credential that was merely
+    # disabled) is still refused in safe mode. For that to work the pattern must
+    # NOT anchor to the physical start of line with ``^\s*``: a single-line
+    # comment marker (``#``, ``;``, ``//``, ``*`` block-continuation, ``dnl``,
+    # ``--``, ``REM`` ...) sits between the line start and ``authorization`` and
+    # would defeat a ``^\s*`` anchor, letting the header slip past the raw scan
+    # (the soft/comment-aware tier never sees it, because that tier strips
+    # comment lines by design). Instead we require the header to be preceded by
+    # start-of-input or any non-identifier character, which is comment-marker
+    # agnostic and still avoids matching an identifier such as
+    # ``my_authorization:``.
     re.compile(
-        rb"(?im)^\s*(?:proxy-)?authorization\s*:\s*"
+        rb"(?im)(?:^|[^A-Za-z0-9_.-])(?:proxy-)?authorization\s*:\s*"
         rb"(?:bearer|basic|token|digest)\s+\S"
     ),
 ]
diff --git a/tests/test_secret_detection.py b/tests/test_secret_detection.py
index 3c98d64..8de887c 100644
--- a/tests/test_secret_detection.py
+++ b/tests/test_secret_detection.py
@@ -30,6 +30,46 @@ def test_authorization_headers_denied():
         ), data
 
 
+def test_commented_authorization_headers_denied():
+    """A commented-out populated Authorization header is a real (disabled)
+    credential and must still be refused in safe mode, regardless of the
+    single-line comment marker used. This is high-confidence material, so it is
+    scanned against the raw bytes; the pattern must not anchor to the physical
+    start of line, or a leading comment marker defeats it (the soft/comment-aware
+    tier never sees it because that tier strips comment lines)."""
+
+    pol = IgnorePolicy(dangerous=False)
+    for data in [
+        b"# Authorization: Bearer eyJhbGciOiJIUzI1NiIs\n",
+        b"#Authorization: Bearer eyJhbGciOiJIUzI1NiIs\n",
+        b"; Authorization: Basic dXNlcjpwYXNzd29yZA==\n",
+        b"// Authorization: Digest abc123\n",
+        b"   //Authorization: Bearer abc123\n",
+        b"* Authorization: Bearer abc123\n",  # block-comment continuation line
+        b"\t# Authorization: Token abc123\n",
+        b"dnl Authorization: Bearer abc123\n",  # m4-style comment
+        b"-- Proxy-Authorization: Basic abc123\n",  # sql/lua-style comment
+        b"REM Authorization: Bearer abc123\n",  # bat-style comment
+        b"# Proxy-Authorization: Bearer xyz\n",
+    ]:
+        assert (
+            pol._content_deny_reason("/etc/svc/app.conf", data) == "sensitive_content"
+        ), data
+
+
+def test_authorization_header_word_not_false_positive():
+    """The Authorization-header pattern must not fire on benign prose mentions,
+    a non-scheme value, or an identifier that merely ends in 'authorization'."""
+
+    pol = IgnorePolicy(dangerous=False)
+    for data in [
+        b"# how to set the authorization header is documented elsewhere\n",
+        b"authorization: required\n",
+        b"my_authorization: bearer x\n",
+    ]:
+        assert pol._content_deny_reason("/etc/svc/app.conf", data) is None, data
+
+
 def test_benign_urls_not_false_positive():
     pol = IgnorePolicy(dangerous=False)
     for data in [