Skip to content

Security review: the Turing Pi BMC firmware fork

Use with mitigations.

The review covers the fork excavador-turing/BMC-Firmware and its installer excavador-turing/BMC-Installer. The use is one: the BMC firmware of the two Turing Pi 2 boards that hold the k3s nodes. A read-only agent cloned the repositories and read the code. It ran nothing from them. Gate 2 of design D3 of the OpenSpec change main-off-the-tailnet asked for this review.

Repository Revision Note
excavador-turing/BMC-Firmware, branch hive 858e64cb439f2077455aa1c58332e14c5c27de37 The commit message names it the v2.44.0 release. 147 commits on top of upstream turing-machines/BMC-Firmware at d9c55b7a, all by one author
excavador-turing/BMC-Installer, branch master cd86ef888e1262e48dc9050f30425d14a4552b2b The revision that tp2bmc/package/bmc_installer/bmc_installer.mk line 8 pins. 5 commits on top of upstream
excavador-turing/bmcd 2027ba5e (v2.40.0) The daemon. 116 commits on top of upstream
excavador-turing/tpi 90c22266 The command line client
BMC-UI v4.1.0 tarball sha256 5c796b65 Matches tp2bmc/package/bmc-ui/bmc-ui.hash and the SHA256SUMS of the release

The firmware is a Buildroot image for the BMC. Its daemon bmcd runs as root on ports 80, 443 and 9110. It can power, reset and flash the four nodes and open their serial consoles. SSH listens on port 22. The image installs from an SD card or from a .tpu package that /sbin/osupdate writes.

The image contacts four endpoints. It reaches api.github.com and github.com for the update check and the install, only when the UI or the API asks (update_check.rs line 94, firmware_catalog.rs line 166). It reaches pool.ntp.org from overlay/etc/chrony.conf line 5. It reaches firmware.turingpi.com only when the operator enables that source. It speaks mDNS on br0. No cron job or timer installs anything. The root crontab holds only mdnsd-guard.

Two attackers are in scope. The first is the fork’s single maintainer, with his GitHub account and CI. The second is anyone on the management network. By default that network includes the four nodes, because the BMC bridge br0 is flat.

1. High, confirmed: one GitHub account is the root of trust of the firmware

Section titled “1. High, confirmed: one GitHub account is the root of trust of the firmware”

tpi-selfupdate (lines 16, 145 to 174 and 504 to 513) resolves the release tag through the GitHub API. It then checks the image against the SHA256SUMS file of the same release. Nobody signs anything. The one-click install of the web UI (bmcd/src/api/legacy.rs lines 535 to 562) runs that script. CI cuts a release on each tag push (.github/workflows/build.yml lines 575 to 592). The organisation lists no public member, and every fork commit has one author. If an attacker takes over that account or its workflow, both BMCs install attacker firmware at the next click. This is not a bug. It is the trust decision that adopting the fork makes.

2. Medium, confirmed: an authenticated caller gets a root shell through the file parameter

Section titled “2. Medium, confirmed: an authenticated caller gets a root shell through the file parameter”

legacy.rs line 1327 builds a path from the file query parameter. streaming_data_service/data_transfer.rs lines 101 to 107 return that whole path as the file name. app/upgrade_worker.rs lines 285 to 310 push that name onto the staging directory. An absolute name replaces the directory. The code opens the result for writing as root and then runs sh -c "osupdate <name>". The type=firmware request goes straight there (legacy.rs lines 1117 to 1150). Nothing checks the characters or the .tpu suffix. A caller with a password, a token or a proxy identity gets a root shell at once, or overwrites /etc/shadow. The caller needs no image and never meets the boot gate. osupdate also deletes the rollback volume on the way. Upstream has the same code (legacy.rs line 626, upgrade_worker.rs lines 180 to 183), so the boards run it today.

3. Medium, confirmed: the factory password works for SSH as root

Section titled “3. Medium, confirmed: the factory password works for SSH as root”

tp2bmc/configs/tp2bmc_defconfig line 9 sets the root password to turing. overlay/etc/ssh/sshd_config sets PermitRootLogin yes and PasswordAuthentication yes. The factory-password gate of bmcd (authentication/factory_password.rs lines 88 to 96) covers /api/bmc only, not SSH. A file factory-reset.txt on an SD card (board/tp2bmc/install.scr) restores the factory password.

/metrics listens on plain HTTP on every address (default_config.yaml lines 3 and 18, main.rs lines 313 to 319) with no authentication. Every host on br0 can read it. It reports the firmware and daemon versions, the node power states, the staged image, the TLS certificate details and the switch counters.

5. Low: a proxy identity can replace the client CA

Section titled “5. Low: a proxy identity can replace the client CA”

api/access.rs lines 345 to 418 let a caller replace the client CA, and nothing refuses a caller whose scheme is mtls. The delete path at lines 425 to 440 has that refusal. A compromised fleet gateway can install its own CA and survive a rotation of the real one.

6. Low: the --tag argument of tpi-selfupdate is not validated

Section titled “6. Low: the --tag argument of tpi-selfupdate is not validated”

Lines 456 to 458 accept any value, and the comment at lines 470 to 471 claims otherwise. The value flows into the download URL and the JSON. Only an authenticated caller reaches it. The install path for an HTTP source is dead code. Nothing calls http_asset, and nothing reads SOURCE_URL after line 433. So an install from firmware.turingpi.com fetches from GitHub and fails.

board/tp2bmc/factory_overlay/upper/etc/ssl/certs/bmcd_key.pem is upstream’s self-signed key, with the name “Turing-Pi self signed” and an expiry in 1970. The factory SD image includes it (post_image.sh lines 27 to 46 and 55, and files_to_delete.txt is empty). The release artifacts do not include it (scripts/build.sh lines 133 to 176). The new certificate generator keeps it when it exists (generate_self_signedx509.sh lines 133 to 138 and 223 to 235). Do not flash the factory image.

  • The nodes share br0 with BMC management (README line 77). The firmware ships VLAN presets in bmcd/src/app/switch_document.rs.
  • The password change feeds chpasswd through stdin (access.rs lines 299 to 311). A newline in a new password applies the first line and reports success.
  • CI grants contents: write only on push and tag, and uses pull_request, not pull_request_target (build.yml lines 22 to 32 and 245 to 248). Every action carries a SHA pin except the one in upstream-watch.yml line 32, which holds issues: write only. The build pulls the toolchain image by a mutable tag (build.yml lines 159 and 373). A workflow dispatch input lands in a shell command (build.yml line 285), and only a caller with write access can trigger it.
  • SysRq is reachable over the serial console with BREAK, which needs physical access (linux_defconfig, S00sysrq).
  • Developer tooling: .envrc evaluates devbox, and Justfile line 32 mounts the parent directory read-only into the build container.
  • No file in the repositories addresses an AI agent. A scan for zero-width and bidirectional characters found nothing. The two HTML comments in the issue templates are benign.
  • The BMC-UI v4.1.0 artifact (658,860 bytes, built by release.yaml from the tag with npm ci) loads no remote script and calls no eval. Its only links are documentation.
  • The RK3588 loader blob bmcd/src/usb_boot/rk3588_spl_loader_v1.08.111.bin (448,960 bytes, sha256 a4765bcc) is identical to upstream.
  • Build the .tpu package from the reviewed revision and install it by hand (README.md lines 141 to 144). scripts/configure.sh line 77 checks Buildroot by sha256, and buildroot_patches/0003 pins the kernel hash. If you install a release instead, read its diff first and prefer --channel stable (tpi-selfupdate lines 44 to 48).
  • Change the root password at first boot. bmcd forces that for the API (factory_password.rs). Install an authorized_keys file and set PasswordAuthentication no in sshd_config on the overlay.
  • Set host: in /etc/bmcd/config.yaml to the management address (default_config.yaml lines 3 and 15 to 17). Keep ports 22, 443 and 9110 away from the node ports with a firewall rule or a VLAN preset.
  • Remove or disable the GitHub and HTTP firmware sources (type=firmware_sources, stored at /mnt/overlay/firmware-sources.json, firmware_sources.rs line 37). That stops the egress and the one-click install.
  • If you use the fleet proxy, pin tls.client_ca in config.yaml. The API then refuses to change it (access.rs lines 352 to 356).
  • upgrade_worker.rs lines 285 to 310 and os_park: take Path::file_name(), reject any character outside [A-Za-z0-9._-], require the .tpu suffix, and run /sbin/osupdate <path> without sh -c.
  • access.rs line 345: add the same mtls refusal that delete_client_ca has.
  • tpi-selfupdate line 456: validate the tag with is_versioned and reject /, " and ...
  • sshd_config: PasswordAuthentication no after the keys are in place.
  • The kernel patches (rtl8370-vlan-support.patch, 928 lines, and the rebased i2c and realtek patches). The reviewer skimmed them. RTL8365MB_VLAN_MAX bounds the VLAN id index, but the reviewer did not read the driver.
  • The UBI preservation parser of BMC-Installer (src/ubi/preserve.rs). Its input needs a prior write to the flash.
  • The 26,000-line delta of bmcd. The reviewer read its authentication, shell calls, file writes, TLS and egress, not every handler, and did not audit the Cargo dependency tree.
  • The BMC-UI bundle. The reviewer did not compare it byte for byte with its source, because the bundle is minified.
  • The mapping from the tag v2.44.0 to commit 858e64cb. It comes from the commit message, not from git tag.