diff --git a/src/dippy/cli/kubectl.py b/src/dippy/cli/kubectl.py
index 3178ac30..b8029532 100644
--- a/src/dippy/cli/kubectl.py
+++ b/src/dippy/cli/kubectl.py
@@ -98,6 +98,98 @@
 }
 
 
+SECRET_RESOURCES = frozenset({"secret", "secrets"})
+
+SAFE_OUTPUT_FORMATS = frozenset({"name", "wide"})
+
+# Flags (after the verb) that consume the next token as a value
+_POST_VERB_FLAGS_WITH_ARG = frozenset(
+    {
+        "-o",
+        "--output",
+        "-n",
+        "--namespace",
+        "-l",
+        "--selector",
+        "-f",
+        "--filename",
+        "--field-selector",
+        "--sort-by",
+        "--template",
+        "--context",
+        "--cluster",
+    }
+)
+
+
+def _is_secret_data_exposure(
+    tokens: list[str],
+    rest: list[str],
+    word_has_expansions: tuple[bool, ...] = (),
+    rest_offset: int = 0,
+) -> bool:
+    """Check if a get command targets secrets with a data-exposing output format.
+
+    Scans rest for the resource type and full tokens for -o (which can appear
+    before or after the verb).  Tokens built from a bash expansion are treated
+    conservatively, since they could resolve to a secret resource or a
+    data-exposing format at runtime.
+    """
+    # Find resource type: first non-flag token in rest
+    resource_type = None
+    resource_abs_pos = None
+    i = 0
+    while i < len(rest):
+        token = rest[i]
+        if token in _POST_VERB_FLAGS_WITH_ARG:
+            i += 2
+            continue
+        if token.startswith("-"):
+            i += 1
+            continue
+        resource_type = token
+        resource_abs_pos = rest_offset + i
+        break
+
+    if resource_type is None:
+        return False
+
+    # If the resource came from an expansion, it could resolve to "secret"
+    if _has_expansion(word_has_expansions, resource_abs_pos):
+        return True
+
+    # Handle comma-separated resources (e.g., "secret,configmap") and
+    # type/name syntax (e.g., "secret/my-secret")
+    parts = resource_type.split(",")
+    if not any(p.split("/")[0] in SECRET_RESOURCES for p in parts):
+        return False
+
+    # Resource IS secrets -- if any remaining token came from an expansion, it
+    # could inject a data-exposing format like -o yaml
+    if _has_expansion_after(word_has_expansions, resource_abs_pos + 1):
+        return True
+
+    # Find output format from full token list (-o can appear before or after verb)
+    output_format = None
+    for j, token in enumerate(tokens):
+        if token in ("-o", "--output") and j + 1 < len(tokens):
+            output_format = tokens[j + 1]
+            break
+        if token.startswith("--output="):
+            output_format = token[len("--output=") :]
+            break
+        if len(token) > 2 and token[:2] == "-o" and token[2] != "-":
+            output_format = token[2:]
+            break
+
+    if output_format is None:
+        return False
+
+    # Extract format name before any = (e.g., "jsonpath='{.data}'" -> "jsonpath")
+    format_name = output_format.split("=")[0]
+    return format_name not in SAFE_OUTPUT_FORMATS
+
+
 def _extract_exec_inner_command(tokens: list[str]) -> list[str] | None:
     """Extract command from kubectl exec args (after -- separator)."""
     try:
@@ -108,9 +200,20 @@ def _extract_exec_inner_command(tokens: list[str]) -> list[str] | None:
         return None  # No -- separator
 
 
+def _has_expansion(word_has_expansions: tuple[bool, ...], pos: int) -> bool:
+    """Whether the token at pos was built from a bash expansion (runtime value)."""
+    return 0 <= pos < len(word_has_expansions) and word_has_expansions[pos]
+
+
+def _has_expansion_after(word_has_expansions: tuple[bool, ...], start: int) -> bool:
+    """Whether any token at or after start was built from a bash expansion."""
+    return any(word_has_expansions[start:])
+
+
 def classify(ctx: HandlerContext) -> Classification:
     """Classify kubectl command."""
     tokens = ctx.tokens
+    expansions = ctx.word_has_expansions
     base = tokens[0] if tokens else "kubectl"
     if len(tokens) < 2:
         return Classification("ask", description=base)
@@ -147,23 +250,42 @@ def classify(ctx: HandlerContext) -> Classification:
         return Classification("ask", description=base)
 
     rest = tokens[action_idx + 1 :] if action_idx + 1 < len(tokens) else []
+    rest_offset = action_idx + 1
     desc = f"{base} {action}"
 
-    # Check for subcommands first
+    # Check for subcommands first (config/auth/rollout)
     if action in SAFE_SUBCOMMANDS and rest:
-        for token in rest:
+        for idx, token in enumerate(rest):
             if not token.startswith("-"):
+                abs_pos = rest_offset + idx
+                if _has_expansion(expansions, abs_pos):
+                    return Classification("ask", description=desc)
                 if token in SAFE_SUBCOMMANDS[action]:
+                    # config view --raw exposes unredacted kubeconfig credentials
+                    if action == "config" and token == "view":
+                        if "--raw" in rest or _has_expansion_after(
+                            expansions, abs_pos + 1
+                        ):
+                            return Classification("ask", description=f"{desc} {token}")
                     return Classification("allow", description=f"{desc} {token}")
                 break
 
     if action in UNSAFE_SUBCOMMANDS and rest:
-        for token in rest:
+        for idx, token in enumerate(rest):
             if not token.startswith("-"):
+                abs_pos = rest_offset + idx
+                if _has_expansion(expansions, abs_pos):
+                    return Classification("ask", description=desc)
                 if token in UNSAFE_SUBCOMMANDS[action]:
                     return Classification("ask", description=f"{desc} {token}")
                 break
 
+    # Sensitive data checks (before blanket safe-action approval)
+    if action == "get" and _is_secret_data_exposure(
+        tokens, rest, expansions, rest_offset
+    ):
+        return Classification("ask", description=f"{desc} (secret data)")
+
     # Simple safe actions
     if action in SAFE_ACTIONS:
         return Classification("allow", description=desc)
diff --git a/tests/cli/test_kubectl.py b/tests/cli/test_kubectl.py
index c19ebecd..6c1263e2 100644
--- a/tests/cli/test_kubectl.py
+++ b/tests/cli/test_kubectl.py
@@ -48,6 +48,7 @@
     ("kubectl get all -A", True),
     ("kubectl get configmaps", True),
     ("kubectl get secrets", True),
+    ("kubectl get secret my-secret", True),  # default table view, no values
     ("kubectl get ingress", True),
     ("kubectl get pv", True),
     ("kubectl get pvc", True),
@@ -205,6 +206,59 @@
     ("kubectl expose deployment nginx --port=80 --target-port=8080", False),
     ("kubectl expose pod nginx --port=80 --type=NodePort", False),
     #
+    # kubectl get secret - sensitive data exposure
+    # Output formats that could expose secret values require confirmation
+    #
+    ("kubectl get secret my-secret -o yaml", False),
+    ("kubectl get secret my-secret -o json", False),
+    ("kubectl get secrets -o yaml", False),
+    ("kubectl get secrets -o json", False),
+    ("kubectl get secret my-secret -o jsonpath='{.data.password}'", False),
+    ("kubectl get secret my-secret -o go-template='{{.data}}'", False),
+    (
+        "kubectl get secret my-secret -o custom-columns=NAME:.metadata.name,DATA:.data",
+        False,
+    ),
+    ("kubectl get secret my-secret --output=yaml", False),
+    ("kubectl get secret my-secret --output json", False),
+    ("kubectl get secret/my-secret -o yaml", False),  # type/name syntax
+    ("kubectl get secret,configmap -o yaml", False),  # comma-separated includes secret
+    ("kubectl -n kube-system get secret my-secret -o yaml", False),  # flags before verb
+    ("kubectl get secret my-secret -o name", True),  # -o name is safe (no values)
+    ("kubectl get secret my-secret -o wide", True),  # -o wide is safe (no values)
+    ("kubectl describe secret my-secret", True),  # describe never shows values
+    ("kubectl get pods -o yaml", True),  # non-secret resource is fine
+    ("kubectl get configmap my-config -o yaml", True),  # non-secret resource is fine
+    #
+    # kubectl get secret - opaque tokens (cmdsubs, param expansions) in arguments
+    #
+    (
+        'kubectl get secret somesecret $(echo "-o yaml")',
+        False,
+    ),  # cmdsub could inject format
+    ("kubectl get secret somesecret `echo '-o yaml'`", False),  # backtick cmdsub
+    ("kubectl get $(echo secret) -o yaml", False),  # cmdsub resource, could be secrets
+    ("kubectl get $RESOURCE -o yaml", False),  # param expansion resource
+    ("kubectl get secret somesecret -o $FORMAT", False),  # param expansion format
+    ("kubectl get secret somesecret -o ${FORMAT}", False),  # braced param expansion
+    ("kubectl get pods $(echo '-o yaml')", True),  # pods aren't secrets, always safe
+    ("kubectl get pods -o $FORMAT", True),  # pods aren't secrets, always safe
+    #
+    # kubectl config view --raw - exposes unredacted kubeconfig credentials
+    #
+    ("kubectl config view --raw", False),
+    ("kubectl config view --raw --minify", False),
+    ('kubectl config view $(echo "--raw")', False),  # cmdsub could be --raw
+    ("kubectl config view $RAW_FLAG", False),  # param expansion could be --raw
+    #
+    # kubectl config/auth/rollout - opaque subcommands
+    #
+    ('kubectl config $(echo "set-context") production', False),  # opaque subcommand
+    ('kubectl describe $(echo "pod nginx")', True),  # describe is always safe
+    ("kubectl describe $RESOURCE", True),  # describe is always safe
+    ('kubectl logs $(echo "nginx")', True),  # logs is always safe
+    ("kubectl logs $POD", True),  # logs is always safe
+    #
     # kubectl exec - delegation to inner command
     #
     ("kubectl exec pod -- cat /etc/passwd", True),  # cat is safe