This document is the v1.0 security assurance record (Wave 7, slice 04). It consolidates self-audit findings from tasks 7.1–7.7, mitigations implemented in tree, and honest scope limits. Read Limitations before treating this as third-party assurance.

Related code: src/sandbox/ · src/gateway/ · src/hardware/ · src/asap/manifest_keys.c

ShellClaw runs as a user-level agent on edge boards (Jetson Orin Nano Super, Raspberry Pi). Untrusted input arrives via chat channels (Telegram, Discord, Web UI) and may invoke:

• Shell tool — subprocess with optional Linux namespaces + allowlist.
• File tool — workspace-scoped paths when configured.
• Hardware tools — GPIO, I2C, camera (gateway /api/hardware/* is Bearer-authenticated in v1.0).
• ASAP / gateway — Bearer-authenticated /api/* HTTP and WebSocket. POST /asap is protocol-public (rate-limited, not Bearer); authenticity is [asap].trusted_senders.

The primary goals are: prevent sandboxed shell commands from escaping to host destruction, block direct GPU/camera daemon access from the shell sandbox, and keep signing keys and cloud credentials off disk with unsafe permissions.

Release gate: zero Critical findings open; Medium items above have documented mitigations and regression tests.

Implementation: src/sandbox/sandbox.c, src/sandbox/allowlist.c.

Audit task 7.1 (2026-05-25). Reviewed src/sandbox/sandbox.c and src/sandbox/allowlist.c against JetPack 6.2.x (kernel 5.15) on Jetson Orin Nano Super.

sandbox_exec() does not call bind() or pivot_root(). After unshare(CLONE_NEWNS) the child makes the copied mount tree MS_REC|MS_PRIVATE and, once the command is PID 1, remounts proc on /proc. It does not bind-mount Tegra GPU devices.

unshare(CLONE_NEWNS | CLONE_NEWNET | CLONE_NEWPID);

In particular, these paths are not explicitly mounted into the shell namespace:

• /dev/nvhost-*
• /dev/nvgpu
• /dev/nvmap

The task checklist references unshare(CLONE_NEWNS) + pivot_root as a hardened pattern. v1.0 does not implement pivot_root. After unshare(CLONE_NEWNS), the child inherits a copy of the host mount tree (made MS_PRIVATE). Jetson /dev nodes remain in that mount namespace; Landlock (when a workspace path is set) is the kernel FS bound and does not grant /dev/nvhost, /dev/nvgpu, or /dev/nvmap.

Mitigation in v1.0: Landlock plus the shell allowlist, which rejects commands whose text references /dev/nvhost, /dev/nvgpu, or /dev/nvmap (substring blocklist). Regression tests live in tests/test_allowlist.c (test_block_jetson_gpu_devices) and tests/test_sandbox.c.

Residual risk: a crafted command that opens GPU nodes without those literal substrings (e.g. shell globs) is still denied by Landlock when a workspace path is set (/dev/nv* is not in the grant list). Without Landlock (non-Linux, or sandbox off) the substring blocklist remains best-effort. workspace_only walks a missing destination’s existing ancestor, collapses .. lexically (without cancelling .. across a symlink), scans quoted/embedded / ~ and relative ../, extracts and percent-decodes file: URLs, fail-closes in-command HOME/PWD assignment, and expands or fail-closes $ on the full command (including glued $IFS and mid-token $HOME). The string scanner is not a language interpreter: open(chr(47)+'etc/passwd') has no path character in the command text and stays residual on the scanner; Landlock is the kernel bound for that class. Encoded-slash / identity-escape cat-and-mouse is frozen. Conservative regex false positives such as awk '/foo/' remain.

src/sandbox/allowlist.c lists Jetson GPU paths (/dev/nvhost, /dev/nvgpu, /dev/nvmap) and Argus socket strings for all boards. These are substring deny rules on the shell command line, not board-specific mounts or runtime board_id checks. On Raspberry Pi or stub builds the entries are inert (no matching devices) but keep one shared allowlist binary. They do not hide /dev nodes from the inherited mount namespace — see GPU device nodes above.

• CLONE_NEWNET — sandboxed shell has no routable network (no interface setup in child).
• CLONE_NEWPID — PID namespace; the isolator forks so the command process is PID 1 (unshare does not move the caller).

Namespace behavior above was verified by source audit. On-device validation on JetPack 6.2.x is a known pending item (not a CI or merge-to-main gate).

hardware_camera_capture() in src/hardware/hardware_camera.c never invokes /bin/sh -c. The backend builds a fixed argv[] and runs fork() + execvp(argv[0], argv) (or a test hook).

GStreamer pipeline elements (nvarguscamerasrc, caps strings, location=) are separate argv entries, not a single interpolated shell string.

Regression tests: tests/test_hardware_camera.c (test_*_injection_rejected, test_no_shell_invocation).

On Jetson CSI cameras, frame capture uses GStreamer nvarguscamerasrc, which is a client of NVIDIA Argus (nvargus-daemon).

ShellClaw does not bind-mount /tmp/argus_socket into the sandboxed shell namespace (sandbox_exec uses unshare without pivot_root, per § Linux sandbox (Jetson)). Landlock does not grant that socket path.

flowchart LR
  LLM[LLM / channels] --> Agent[ShellClaw agent process]
  Agent -->|camera_capture tool| Cam[hardware_camera_capture]
  Cam -->|fork execvp| GST[gst-launch-1.0]
  GST -->|nvarguscamerasrc plugin| Socket["/tmp/argus_socket"]
  Socket --> Daemon[nvargus-daemon root]
  Agent -->|shell tool sandbox on| SB[sh -c in namespaces]
  SB -.->|blocked allowlist| Socket

• Allowed (by design): the main agent process invokes hardware_camera_capture() → execvp("gst-launch-1.0", …) with nvarguscamerasrc in argv. The GStreamer child connects to Argus as any normal camera client would.
• Blocked: the shell tool path runs inside sandbox_exec() + allowlist. Commands that reference /tmp/argus_socket or argus_socket are rejected (src/sandbox/allowlist.c). Regression: tests/test_allowlist.c (test_block_argus_socket).

• Gateway: POST /api/hardware/camera/snapshot returns a v1.2 placeholder (no HTTP-driven capture in v1.0).
• LLM tool: camera_capture remains registered when the camera backend is active; it uses the same Argus path above and is subject to channel allowlists / operator trust, not the shell sandbox.

Residual risk: a compromised agent process (not merely a sandboxed shell command) could still spawn gst-launch or other Argus clients. Mitigation is process-level trust and minimizing agent attack surface; socket blocking applies to the shell sandbox boundary only.

Audit (2026-05-25): src/gateway/http_lws.c requires_auth() treats every /api/ path as Bearer-protected (except /health, /pair, /.well-known/, /). Unauthenticated requests receive 401 before dispatch_route() runs.

Handlers: src/gateway/routes_hardware.c. Non-GET methods on read-only paths return 405. Example: POST /api/hardware/board → 405 (tested in test_routes_hardware).

POST /api/hardware/camera/snapshot returns the v1.2 deferred JSON stub when authenticated. Per-token throttling was removed from src/gateway/rate_limit.c until the real HTTP image path ships in Phase 7; Bearer auth still applies via requires_auth().

Slice 02 introduced these routes; this audit confirms Bearer gating remains centralized in http_lws.c. Rate limiting for camera POST is deferred with the capture implementation.

POST /asap is excluded from Bearer pairing so ASAP peers can call without a ShellClaw token. Authenticity is [asap].trusted_senders. An empty list (local/dev default) allows every claimed URN.

After provider/tool wiring in handle_asap (#53), a sender that passes that check can:

• task.request — agent_run() with the same tool table as chat (shell, file, hardware, asap_invoke, …)
• mcp.tool_call — execute() with attacker-chosen name and arguments (no LLM)

Default gateway bind is 127.0.0.1 (http_start sets libwebsockets info.iface from gateway.host), which contains this for stock installs. Bind-all hosts (0.0.0.0, *, ::, [::], empty) require allow_bind_all. Production MUST set trusted_senders before exposing the gateway (allow_bind_all, tunnel, marketplace URL). Restricting the inbound MCP tool table (or failing closed when the allowlist is empty and the host is not loopback) is a follow-up.

Audit (2026-05-26): Re-verified slice 03 key handling in src/asap/manifest_keys.c. Keys load lazily on the first GET /.well-known/asap/manifest.json via manifest_keys_ensure_loaded() in src/gateway/routes.c; CLI/agent without gateway does not require ~/.shellclaw/keys/ at startup.

SHELLCLAW_HOME overrides the base directory (same as config and auth tokens).

manifest_priv_permissions_ok() rejects ed25519.priv when any group or other permission bit is set (st_mode & 0077). This covers 0644, 0660, and world-readable modes. New and rotated keys are written with open(…, 0600) plus fchmod(…, 0600) after write.

The keypair is loaded (or created) on the first signed manifest request, not at init_subsystems(). A world-readable ed25519.priv blocks manifest discovery but does not block one-shot -m runs that never hit the gateway manifest route.

• tests/test_manifest_keys: test_manifest_keys_first_run_creates (0600 on create), test_manifest_keys_rejects_loose_priv_perms, test_manifest_keys_rotate_rejects_loose_priv
• tests/test_bootstrap_keys: CLI starts without keys; manifest_keys_ensure_loaded rejects loose priv
• tests/test_daemon_smoke.sh: post---rotate-keys stat expects mode 600 on both key files

Residual note: Only ed25519.priv permissions are checked on load; ed25519.pub is not re-stat'd for mode (public material). Owner-only 0400 on the private key passes the group/other check but is not exactly 0600; created keys always use 0600.

Audit (2026-05-26): Re-ran make static and make test-sanitize on the development branch; both exit 0. CI runs the same steps on ubuntu-24.04 (.github/workflows/ci.yml).

Curated suppressions in the Makefile cover known false positives (e.g. knownConditionTrueFalse in config reload paths, TweetNaCl variableScope). This is stricter than a one-off --enable=all run, which floods style noise on cJSON and ASAP helpers without improving security signal.

Local: make static

Local (matches CI):

make test-sanitize

Full pre-release mirror on Linux: ./scripts/ci-local.sh (static → test → LIBGPIOD=0 test → sanitizer → release size → coverage).

Note: Sanitizer builds are slower and require a toolchain with ASan/UBSan support (GCC/Clang on Linux or macOS). They are not used for release binaries (make release).

This section summarizes Jetson Orin Nano Super / JetPack 6.2.x concerns that do not apply to Raspberry Pi or generic x86 dev machines.

v1.0 scope on Jetson: HTTP camera snapshot and sensor decoders are deferred stubs; LLM camera_capture and GPIO/I2C tools remain operator-trusted paths outside the shell sandbox. See NVIDIA Argus and Gateway hardware API.

Validation outside CI: gpio-mockup local ritual plus the on-device checklist in JETSON_SIGNOFF.md. Jetson sign-off is known pending (no Jetson runner in GitHub Actions; not a merge-to-main gate).

Automated regression coverage: tests/test_allowlist.c, tests/test_sandbox.c, tests/test_hardware_camera.c, tests/test_gateway_http.c, tests/test_rate_limit.c, tests/test_manifest_build, tests/test_manifest_keys.

src/sandbox/allowlist.c lists Jetson GPU device path substrings (/dev/nvhost, /dev/nvgpu, /dev/nvmap) and Argus socket paths for all boards. These are deny-by-path rules in the generic shell blocklist, not runtime board_id_t guards: on non-Jetson hosts the entries are harmless no-ops; on Jetson they block sandboxed shell access to GPU nodes and the Argus IPC socket. Camera capture via hardware_camera_capture() runs outside the shell sandbox by design.

Per project decision Q-SECREVIEW (DR-020):

• Self-audit only for v1.0. This document, source review notes in tasks 7.1–7.6, and CI static/sanitizer runs constitute the full assurance package for the v1.0 release. There was no external penetration test, no third-party code review, and no paid security auditor for this version.
• Bug bounty: not offered for v1.0; considered post-v1.0 (likely v1.2+ alongside external review — see roadmap).
• Hardware CI gap: GitHub Actions does not execute on a physical Jetson. Namespace and libgpiod behavior on Tegra are validated via source audit, unit tests, and the gpio-mockup release ritual. The on-device sign-off checklist is known pending and is not a merge-to-main gate.
• Sandbox depth: substring blocklists and namespace isolation are defense in depth, not a formal proof against a determined attacker with shell access when sandbox mode is disabled or when the agent process itself is compromised.

Do not describe ShellClaw v1.0 as "externally audited" or "pen-tested."

Last updated: 2026-05-26 — Wave 7.7 published (tasks 7.1–7.7); satisfies doc task 9.4.