diff --git a/hscontrol/policy/policy_test.go b/hscontrol/policy/policy_test.go
index 65f47b4a..b62e94db 100644
--- a/hscontrol/policy/policy_test.go
+++ b/hscontrol/policy/policy_test.go
@@ -891,9 +891,11 @@ func TestReduceNodesFromPolicy(t *testing.T) {
   ]
 }`,
 			node: n(1, "100.64.0.1", "mobile", "mobile"),
+			// autogroup:internet does not generate packet filters - it's handled
+			// by exit node routing via AllowedIPs, not by packet filtering.
+			// Only server is visible through the mobile -> server:80 rule.
 			want: types.Nodes{
 				n(2, "100.64.0.2", "server", "server"),
-				n(3, "100.64.0.3", "exit", "server", "0.0.0.0/0", "::/0"),
 			},
 			wantMatchers: 1,
 		},
diff --git a/hscontrol/policy/policyutil/reduce_test.go b/hscontrol/policy/policyutil/reduce_test.go
index 09342db7..0851e303 100644
--- a/hscontrol/policy/policyutil/reduce_test.go
+++ b/hscontrol/policy/policyutil/reduce_test.go
@@ -353,10 +353,12 @@ func TestReduceFilterRules(t *testing.T) {
 				},
 			},
 			want: []tailcfg.FilterRule{
-				// Merged: Both ACL rules combined (same SrcIPs and IPProto)
+				// Only the internal:* rule generates filters.
+				// autogroup:internet does NOT generate packet filters - it's handled
+				// by exit node routing via AllowedIPs, not by packet filtering.
 				{
 					SrcIPs: []string{"100.64.0.1/32", "100.64.0.2/32", "fd7a:115c:a1e0::1/128", "fd7a:115c:a1e0::2/128"},
-					DstPorts: append([]tailcfg.NetPortRange{
+					DstPorts: []tailcfg.NetPortRange{
 						{
 							IP:    "100.64.0.100/32",
 							Ports: tailcfg.PortRangeAny,
@@ -365,7 +367,7 @@ func TestReduceFilterRules(t *testing.T) {
 							IP:    "fd7a:115c:a1e0::100/128",
 							Ports: tailcfg.PortRangeAny,
 						},
-					}, hsExitNodeDestForTest...),
+					},
 					IPProto: []int{v2.ProtocolTCP, v2.ProtocolUDP, v2.ProtocolICMP, v2.ProtocolIPv6ICMP},
 				},
 			},
diff --git a/hscontrol/policy/v2/filter.go b/hscontrol/policy/v2/filter.go
index ae01fe2f..71540ea8 100644
--- a/hscontrol/policy/v2/filter.go
+++ b/hscontrol/policy/v2/filter.go
@@ -61,6 +61,12 @@ func (pol *Policy) compileFilterRules(
 				continue
 			}
 
+			// autogroup:internet does not generate packet filters - it's handled
+			// by exit node routing via AllowedIPs, not by packet filtering.
+			if ag, isAutoGroup := dest.Alias.(*AutoGroup); isAutoGroup && ag.Is(AutoGroupInternet) {
+				continue
+			}
+
 			ips, err := dest.Resolve(pol, users, nodes)
 			if err != nil {
 				log.Trace().Caller().Err(err).Msgf("resolving destination ips")
@@ -259,6 +265,12 @@ func (pol *Policy) compileACLWithAutogroupSelf(
 					continue
 				}
 
+				// autogroup:internet does not generate packet filters - it's handled
+				// by exit node routing via AllowedIPs, not by packet filtering.
+				if ag, isAutoGroup := dest.Alias.(*AutoGroup); isAutoGroup && ag.Is(AutoGroupInternet) {
+					continue
+				}
+
 				ips, err := dest.Resolve(pol, users, nodes)
 				if err != nil {
 					log.Trace().Caller().Err(err).Msgf("resolving destination ips")
diff --git a/hscontrol/policy/v2/tailscale_routes_compat_test.go b/hscontrol/policy/v2/tailscale_routes_compat_test.go
index 247a9738..d7cfbb52 100644
--- a/hscontrol/policy/v2/tailscale_routes_compat_test.go
+++ b/hscontrol/policy/v2/tailscale_routes_compat_test.go
@@ -1124,31 +1124,15 @@ func TestTailscaleRoutesCompatExitNodes(t *testing.T) {
 				"user1":         wildcardFilter,
 			},
 		},
-		// TODO: Fix autogroup:internet to not generate filters
-		//
 		// B8: autogroup:internet generates no filters
 		//
-		// TAILSCALE BEHAVIOR:
-		// - autogroup:internet is handled by exit node routing, not packet filters
-		// - ALL nodes should get null/empty filters for autogroup:internet destination
-		// - Traffic is routed through exit nodes via AllowedIPs, not filtered
-		//
-		// HEADSCALE BEHAVIOR:
-		// - Exit nodes (exit-node, multi-router) incorrectly receive filters
-		// - Because 0.0.0.0/0 "covers" autogroup:internet destinations
-		//
-		// ROOT CAUSE:
-		// Headscale treats autogroup:internet like a regular destination and
-		// distributes filters to nodes whose routes cover it (exit nodes)
-		//
-		// FIX REQUIRED:
-		// autogroup:internet should never generate packet filters
+		// autogroup:internet is handled by exit node routing via AllowedIPs,
+		// not by packet filtering. ALL nodes should get null/empty filters.
 		{
 			name: "B8_autogroup_internet_no_filters",
 			policy: makeRoutesPolicy(`
 		{"action": "accept", "src": ["autogroup:member"], "dst": ["autogroup:internet:*"]}
 	`),
-			/* EXPECTED (Tailscale):
 			wantFilters: map[string][]tailcfg.FilterRule{
 				"client1":       nil,
 				"client2":       nil,
@@ -1160,26 +1144,6 @@ func TestTailscaleRoutesCompatExitNodes(t *testing.T) {
 				"big-router":    nil,
 				"user1":         nil,
 			},
-			*/
-			// ACTUAL (Headscale):
-			// Non-exit nodes correctly get nil.
-			// INCORRECT: exit-node and multi-router get filters with expanded public
-			// CIDR ranges from util.TheInternet() (all public IPs excluding CGNAT,
-			// private ranges, and Tailscale ULA). The exact CIDRs are complex and
-			// impractical to list here. Skipping comparison for exit nodes but
-			// documenting the difference: Tailscale returns nil, Headscale returns
-			// filters with SrcIPs=member IPs and DstPorts=expanded public CIDRs.
-			wantFilters: map[string][]tailcfg.FilterRule{
-				"client1":       nil,
-				"client2":       nil,
-				"subnet-router": nil,
-				"ha-router1":    nil,
-				"ha-router2":    nil,
-				"big-router":    nil,
-				"user1":         nil,
-				// exit-node and multi-router omitted - they incorrectly receive filters
-				// with expanded autogroup:internet CIDRs in Headscale (Tailscale: nil)
-			},
 		},
 		// B3: Exit node advertises exit routes (verify RoutableIPs)
 		//