fix(supervisor): drop sandbox child capability bounding set

architecture/sandbox.md

@@ -14,7 +14,12 @@ Each sandbox workload has two trust levels:
 | Agent child | Runs as an unprivileged user with filesystem, process, and network restrictions applied. |
 
 The supervisor keeps enough privilege to manage the sandbox, but the agent child
-loses that privilege before user code runs.
+loses that privilege before user code runs. On Linux, child setup clears the
+capability bounding set during privilege drop so later execs cannot regain
+container-granted capabilities. This is fail-closed: the supervisor retains
+`CAP_SETPCAP` solely to perform the clear, and spawning the workload or SSH shell
+aborts unless the bounding set ends up empty. A `setpcap` `EPERM` is tolerated
+only when the set is already empty; any other outcome fails the spawn.
 
 ## Startup Flow
 

crates/openshell-driver-podman/README.md

@@ -42,7 +42,7 @@ The container spec in `container.rs` sets these security-critical fields:
 |---|---|---|
 | `user` | `0:0` | The supervisor needs root inside the container for namespace creation, proxy setup, Landlock, seccomp, and filesystem preparation. |
 | `cap_drop` | Selected unneeded defaults | Podman's default capability set is already restricted. The driver drops capabilities the supervisor does not need. |
-| `cap_add` | `SYS_ADMIN`, `NET_ADMIN`, `SYS_PTRACE`, `SYSLOG`, `DAC_READ_SEARCH` | Grants supervisor-only capabilities required for namespace setup, process identity, and bypass diagnostics. |
+| `cap_add` | `SYS_ADMIN`, `NET_ADMIN`, `SYS_PTRACE`, `SYSLOG`, `DAC_READ_SEARCH`, `SETPCAP` | Grants supervisor-only capabilities required for namespace setup, process identity, bypass diagnostics, and child bounding-set cleanup. |
 | `no_new_privileges` | `true` | Prevents privilege escalation after exec. |
 | `seccomp_profile_path` | `unconfined` | The supervisor installs its own policy-aware BPF filter. A container-level profile can block Landlock/seccomp syscalls during setup. |
 | `mounts` | Private tmpfs at `/run/netns` | Lets the supervisor create named network namespaces in rootless Podman. |
@@ -98,12 +98,15 @@ openshell sandbox create \
 | `SYS_PTRACE` | Reading `/proc/<pid>/exe` and walking process ancestry for binary identity. |
 | `SYSLOG` | Reading `/dev/kmsg` for bypass-detection diagnostics. |
 | `DAC_READ_SEARCH` | Reading `/proc/<pid>/fd/` across UIDs so the proxy can resolve the binary responsible for a connection. |
+| `SETPCAP` | Clearing the restricted child process capability bounding set before exec. |
 
 The driver intentionally keeps Podman's default `SETUID`, `SETGID`, `CHOWN`,
 and `FOWNER` capabilities because the supervisor needs them to drop privileges
-and prepare writable sandbox directories. It drops unneeded defaults such as
+and prepare writable sandbox directories. It also keeps `SETPCAP` until child
+setup so `drop_privileges()` can clear the child capability bounding set before
+exec. It drops unneeded defaults such as
 `DAC_OVERRIDE`, `FSETID`, `KILL`, `NET_BIND_SERVICE`, `NET_RAW`, `SETFCAP`,
-`SETPCAP`, and `SYS_CHROOT`.
+and `SYS_CHROOT`.
 
 ## Supervisor Sideloading
 

crates/openshell-driver-podman/src/container.rs

@@ -877,8 +877,6 @@ pub fn build_container_spec_with_token_and_gpu_devices(
             "NET_RAW".into(),
             // Not needed: the supervisor does not manipulate file capabilities.
             "SETFCAP".into(),
-            // Not needed: the supervisor does not manage its own capability bounding set.
-            "SETPCAP".into(),
             // Not needed: the supervisor does not call chroot().
             "SYS_CHROOT".into(),
         ],
@@ -899,13 +897,18 @@ pub fn build_container_spec_with_token_and_gpu_devices(
             // Without it the proxy cannot determine which binary made each outbound
             // connection and all traffic is denied.
             "DAC_READ_SEARCH".into(),
+            // Child setup clears the capability bounding set before exec, which
+            // requires CAP_SETPCAP in the supervisor until drop_privileges().
+            "SETPCAP".into(),
         ],
-        // SETUID, SETGID, CHOWN, and FOWNER are intentionally kept from Podman's
-        // default set and not dropped:
+        // SETUID, SETGID, SETPCAP, CHOWN, and FOWNER are intentionally kept from
+        // Podman's default set and not dropped:
         //   SETUID/SETGID – drop_privileges(): setuid()/setgid()/initgroups() to the
         //                   sandbox user. In rootless Podman cap_drop:ALL removes them
         //                   from the bounding set even though uid=0 owns the user
         //                   namespace — so we keep them by not dropping them explicitly.
+        //   SETPCAP       – drop_privileges(): clears the child capability
+        //                   bounding set before the sandbox user execs.
         //   CHOWN         – prepare_filesystem(): chown(path, uid, gid) on newly
         //                   created read_write directories so the sandbox user can
         //                   write to them.
@@ -1451,12 +1454,14 @@ mod tests {
             added.contains(&"DAC_READ_SEARCH"),
             "missing DAC_READ_SEARCH"
         );
+        assert!(added.contains(&"SETPCAP"), "missing SETPCAP");
 
         // SETUID and SETGID are NOT in cap_add — they remain available from the
         // default bounding set because we no longer use cap_drop:ALL. Verify they
-        // are also not explicitly dropped. Similarly CHOWN and FOWNER must not be
-        // dropped because prepare_filesystem() calls chown() on newly created
-        // read_write directories before the supervisor drops privileges.
+        // are also not explicitly dropped. Similarly SETPCAP, CHOWN and FOWNER
+        // must not be dropped because child setup clears the bounding set and
+        // prepare_filesystem() calls chown() on newly created read_write
+        // directories before the supervisor drops privileges.
         let dropped: Vec<&str> = spec["cap_drop"]
             .as_array()
             .expect("cap_drop should be an array")
@@ -1473,6 +1478,10 @@ mod tests {
             !dropped.contains(&"FOWNER"),
             "FOWNER must not be dropped (needed for chown on non-owned files)"
         );
+        assert!(
+            !dropped.contains(&"SETPCAP"),
+            "SETPCAP must not be dropped (needed for child bounding-set clear)"
+        );
         assert!(
             !dropped.contains(&"ALL"),
             "must not use cap_drop:ALL in rootless Podman"

crates/openshell-supervisor-process/Cargo.toml

@@ -35,6 +35,7 @@ libc = "0.2"
 rustix = { workspace = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
+capctl = "0.2.4"
 landlock = "0.4"
 seccompiler = "0.5"
 tempfile = "3"