TheMaxMur/RS-Key v0.4.5

RS-Key v0.4.5

on GitHub

[0.4.5] - 2026-08-03

Security

Audit run-31 fixes (bcdDevice 0x085F, rsk 0.3.23, rsk-tui 0.3.3):

• A cancel can no longer end another transport's touch ceremony.
  CANCEL_REQUESTED was one global flag, so an unprivileged process holding only
  the FIDO HID nub could CTAPHID_CANCEL an OpenPGP, PIV, OATH or Yubico-OTP
  ceremony — WORKER_LOCK does not serialize the two, because a parked FIDO
  request acquires it inside the future the keepalive loop is already driving.
  On a screenless build the next ceremony then started with spent == false, so
  the user's descending finger could confirm the attacker's makeCredential /
  getAssertion instead. The single flag becomes a typed WAIT_SCOPE set around
  every dispatch; request_cancel honours it exactly as cancel_otp_wait already
  did, and the keepalive advertises UPNEEDED only for the channel that owns the
  wait (which also closes a cross-transport "a touch is imminent" oracle).
• The trusted display's device PIN is a first-class credential on the host path.
  The FIDO vendor gate keyed solely on the clientPIN, so a display build whose owner
  completed the panel's own onboarding — device PIN set, no clientPIN — exported its
  master seed on one touch, and a panel lock did not stand in the way (a host
  ceremony paints over it). pin_gate now falls back to a device-PIN entry on the
  device's own pad, covering BACKUP_EXPORT, BACKUP_LOAD, ATT_IMPORT,
  ATT_CLEAR and the audit subcommands.
• The on-device auto-lock has its own deadline. It rode on the display-sleep
  timer, which every host ceremony refreshed — including the ungated
  authenticatorSelection — so a loop of them held the panel unlocked for the whole
  plugged-in session. The lock now counts from the last local interaction, and it
  re-arms without blanking, so "Display sleep: Off" no longer switches a security
  control off with it.
• Seal backup and Firmware → reboot-to-BOOTSEL take the device PIN, like every
  other irreversible panel action. Sealing cannot be undone except by a factory
  reset that destroys the seed it protects, and BOOTSEL is the entry point for the
  issue-#37 flash-rollback.
• The passkey rename is device-PIN gated, and the delete card names the real
  relying party. A nickname replaces the rpId on the browse screens, so an
  unauthenticated relabel could aim the owner's own PIN-gated delete at the wrong
  credential.
• A host can no longer hold the on-device UI shut. Modals abandoned themselves
  the instant a host command queued, with no floor — and because REQ latches until
  the worker drains it, one repeated ungated command kept the unlock pad closing on
  its first poll. Entry and hold modals now keep a short guaranteed slice and never
  yield mid-entry or mid-hold.
• Touch, the wake button and the auto-lock no longer depend on USB being
  configured. They sat behind the LED status, which stays at Boot until a host
  completes SET_CONFIGURATION, so on charger or battery power the panel animated
  but ignored every touch.
• The at-rest scrub is re-armed by the lazy OpenPGP migration. The one-shot
  compaction lap latched at boot, but the OpenPGP DEK chain and its verifiers migrate
  off the pre-OTP key base on the first PIN verify — appends, leaving the superseded
  chip-serial-rooted copies readable in a flash dump forever. That contradicted the
  threat model's "a flash dump cannot brute-force the PIN offline": the pre-OTP root
  derives from the public chip serial. EF_HARDENED moves to rsk-fs and the
  migration clears it, so the next boot scrubs.
• A host-requested warm reboot no longer advances the Yubico-OTP use counter.
  power_up_bump ran on every main, so ~32768 ungated warm reboots saturated the
  15-bit counter while the RAM session counter restarted at 0 — leaving the key
  re-emitting (useCtr, sessionCtr) pairs a validation server rejects as replays.
• The OTP keyboard transport zeroizes its buffers (frame reassembly, the taken
  request, the type queue) and joins the pre-BOOTSEL scrub. They could hold a slot's
  AES key, private UID, access code or static password.
• forcesig holds on OpenPGP PSO:CDS. PW3 is still accepted for parity, but not
  when the card is configured "PW1 valid for one signature" — only PW1 can be cleared
  per signature, so an admin-PIN entry would otherwise have authorised unlimited
  signatures silently.
• TERMINATE DF fails instead of reporting a wipe it could not prove.
  wipe_openpgp used delete (which skips a false-absent file for_each_key keeps
  yielding, so the sweep could spin) and discarded the enumeration's completeness
  flag. It now matches the FIDO and PIV sweeps: force_delete, a delete budget, and
  MEMORY_FAILURE on a truncated walk.
• rsk offboard, rsk inventory verify and rsk backup restore/finalize refuse
  to guess between two attached keys. The PC/SC half picks its device by reader
  name and the HID half took the first match, so offboard could confirm one device's
  serial and factory-reset another's FIDO identity, and inventory verify could bind
  a serial to a different device's attestation key — the enrollment anchor
  docs/guides/fleet.md tells operators to record. ctaphid.find_all is new;
  connect_fido(exclusive=True) gates the destructive callers.
• rsk backup export no longer prints the PIN beside the mnemonic. The export is
  gated on touch plus the PIN; echoing it into the block the user was just told to
  record collapsed both factors into one artifact.
• rsk-tui: no first-reader fallback, a real reboot status word, and the audit
  window cross-check. With CCID absent the cockpit connected to whatever card was
  in a reader — SELECTing five applets on it every 5 s and rendering its PIN counters
  as the RS-Key's. reboot reported a declined on-screen gate as success. audit_read
  kept only the modulo check under a comment claiming parity with audit.py, so a
  device could present 1 of N events as a complete window. The header also refuses to
  vouch for an identity when more than one FIDO device is attached.

Fixed

• Board files: flash size and LED pins. waveshare-touch-lcd and tenstar-usb
  declared size_mb = 4 where every other reference (nix, CI, the docs, the flash-map
  diagram) builds them at 16M — a 12 MiB shift of the key store, which reads as an
  empty device and boots a display build unlocked. tenstar-usb and seeed-xiao
  pointed the WS2812 at GP16 (the Waveshare One's pin) instead of the hardware-verified
  GP22, leaving the consent indicator dark. Each shipped board file is now smoke-built
  in CI, and BOARD= is documented in docs/build.md.
• build.rs rejects instead of truncating. u8() wrapped an out-of-range pin into
  a plausible one before the resolvers' range asserts could see it, and the four
  display control pins had no range check at all — a value ≥ 128 aliases onto a real
  GPIO through embassy's bit-7-banked AnyPin. Board-file booleans now accept the same
  spellings the env resolvers do (yes/on) and panic on anything else, rather than
  silently reading as false; presence.source is matched case-insensitively.
• rsk backup finalize works on a device with a PIN set (rsk 0.3.22). It
  sent the vendor command with no pinUvAuthToken, so pin_gate answered
  CTAP2_ERR_PUAT_REQUIRED (0x36) and the one-time export window could not be
  sealed once clientPin was enabled; it now takes the same _gated() path
  rsk backup export already used
  (#59, thanks
  @lockedmutex).
• A touch-gated challenge no longer looks like a timeout the moment the button
  is pressed. While a command ran, the keyboard transport picked its status
  byte from the live presence flag — 0x20 while a touch was awaited, 0x10
  otherwise. But that flag drops as soon as the press is collected, and the
  response only appears once the HMAC has been computed, so every touch-gated
  challenge served a short 0x10 window in between: measured at 9 ms against
  Windows and 11 ms against macOS. ykpers' blocking read
  (yk_wait_for_key_status, which KeePassXC vendors) arms itself on 0x20 and
  then reads any byte carrying neither the pending nor the waiting bit as "the
  key timed out waiting for the user" — so a host polling inside that window
  abandoned a challenge the key had already answered. A YubiKey never shows it:
  it reports the wait, plus a seconds countdown, right up to the response. The
  wait now latches for the rest of the command, so only the response — or the
  idle status frame after a real timeout — replaces it, which leaves an expired
  wait ending exactly as promptly as before. bcdDevice → 0x085C.
• The PIN-entry band no longer leaves a stale "+" overflow marker and the right
  half of the 10th dot behind when the user deletes from a long PIN back to ≤10
  digits (render_pin_dots). The repaint cleared one small rectangle per dot,
  centred on each circle, but masked_entry draws dots top-left-aligned — so the
  clear was off by ENTRY_DIA/2 and missed the "+" at x 184 plus dot 10's right
  tail (x 176..180). Deleting 11 → 10 left the "+" on screen for the rest of the
  session, lying that the PIN was still long, and 10 → 9 left a one-pixel stub.
  The repaint now clears one strip over the whole entry band (every dot position
  plus the overflow slot to its right) before redrawing. Covered by a host test
  in rsk-ui.

Audit run-30 fixes (bcdDevice 0x085D, rsk 0.3.21, rsk-tui 0.3.2):

• The OTP frame protocol is served on the keyboard interface only again. It
  had also been answered on the FIDO HID interface to match a YubiKey; on macOS
  that removed a privilege boundary — IOKit gates a keyboard-usage HID nub behind
  Input Monitoring while the 0xF1D0 FIDO nub opens to any console-user process,
  so slot programming and challenge-response became reachable unprompted. The
  keyboard interface is already index 0 (what index-addressing hosts need), so the
  FIDO door gained nothing and is removed.
• rsk openpgp reset no longer risks destroying an unrelated OpenPGP card. It
  now checks the SELECT status (refusing a card with no OpenPGP application) and
  takes the typed confirmation the docs already promised, and find_reader fails
  with a clear error instead of silently grabbing the first PC/SC reader when no
  RS-Key is present.
• A slot UPDATE no longer resets the Yubico-OTP use counter / OATH-HOTP moving
  factor. It built a 52-byte record, dropping the 8-byte counter tail, so a
  routine ykman otp settings silently rolled the anti-replay counter back. The
  tail is now carried forward; only a full re-CONFIGURE resets it.
• A YubiKey config-lock code is no longer stored or disclosed. RS-Key does not
  implement the lock, but WRITE CONFIG kept the 16-byte code and READ CONFIG
  echoed it in cleartext to any unauthenticated host. The lock tags are now
  stripped on write, and READ CONFIG always reports the lock unset (as hardware
  does).
• rsk offboard always writes a receipt now. A missed touch (or a malformed
  device) on the post-wipe journal read used to abort before the receipt was
  written, leaving an irreversible wipe with no artifact; the failure now degrades
  to a note in a written receipt.
• rsk backup status sanitizes device output. The sealed/has_seed values
  from the device are coerced to booleans, so a hostile device can no longer inject
  terminal escapes during the seed-backup ceremony.
• rsk-tui no longer prints "identity verified ✓" for an unpinned device. The
  verifying key comes from the same response being checked, so a self-signed
  counterfeit passed; the verdict now states only that the signature is
  self-consistent and points at rsk inventory verify --expect-key, matching the
  CLI.
• The host tools bound CTAPHID response reassembly by wall clock. A device
  trickling one small continuation frame per timeout could hang rsk or freeze
  the TUI for hours; both now enforce the same 120 s deadline the keepalive loop
  uses.
• The release workflow refuses a tag that is not an ancestor of main as
  defence in depth. The primary control against a leaked write token laundering
  unreviewed code into a signed release is a repository tag ruleset restricting
  who may create refs/tags/v* — configure that in the repo settings.
• docs/unsafe.md records the four new AnyPin::steal sites (display
  CS/DC/RST/TP_RST) and the firmware/build.rs env::set_var site,
  restoring the runtime-site count from 15 to 19 and matching the new
  collision-assert containment in the prose.

Added

• Per-board build configuration: BOARD=<name> picks firmware/boards/<name>.toml
  instead of setting the LED / presence / flash / display env vars one by one.
  Ships with waveshare-one, waveshare-touch-lcd, tenstar-usb, and
  seeed-xiao. The individual env vars still work and still win over the board
  file, so existing build recipes are unaffected. The display's SPI1
  (PIN_10/11/12) and I2C1 (PIN_6/7) lines stay hard-wired; only the control
  GPIOs (cs/dc/rst/tp_rst/backlight), bus frequencies, colour inversion,
  and colour order are per-board.
• Trusted-display UI redesign (--features display): anti-aliased circles for
  the boot, reset, and PIN-entry dots; a shared component system (card,
  rect_card, list::group_card, list::row) behind the Passkeys, Audit,
  Applets, and Settings screens; and a unified 10 px gap between the title bar and
  the content on every screen. The passkey rename screen replaces the up/down
  character wheel with a T9 phone-style keypad — a repeated press cycles the
  key's letter group, a different key or an 800 ms pause commits, and the field
  and keypad repaint in place instead of clearing the frame.

Changed

• bcdDevice bumped to 0x085E for the UI redesign and the per-board
  configuration system.
• OpenPGP RSA heap restored to 128 KiB. It was halved to 64 KiB inside the
  per-board-config commit without a callout; on embedded_alloc a failed
  allocation aborts (handle_alloc_error → panic → watchdog reset), so a long
  RSA-4096 keygen/CRT mid-operation could reset the device. Back to the v0.4.4
  value until a separate justification for the smaller size is on record.
• Display control GPIOs are now checked for collisions at compile time. The
  four new AnyPin::steal sites for CS/DC/RST/TP_RST were added by number
  with no compile-time guard beyond WAKE_PIN vs the 10..=18 range, so a board
  config could aim cs at the same pad as WAKE_PIN, LED_PIN, the hard-wired
  SPI1 (PIN_10/11/12) / I2C1 (PIN_6/7) lines, or another control line, and
  silently drive one pad from two owners. A const _: () = assert!(...) now
  rejects all of those at build time. The backlight PWM (pin, slice, channel)
  combo likewise had no compile-time guard and panicked at boot on an unsupported
  board (a runtime panic on fully constant operands); it too is now a const
  assert, and the runtime match arm is unreachable!.
• firmware/build.rs re-rustc's when any of the 11 PK_DISPLAY_* env vars
  change. They were the only env knobs missing a cargo:rerun-if-env-changed,
  so overriding PK_DISPLAY_CS etc. without touching BOARD reused the cached
  build and shipped a firmware with stale pins.

Removed

• Dead code from the UI redesign: aa::filled_rounded_rect (~64 lines, never
  called), and the CARET_BLINK_MS const left behind with #[allow(dead_code)]
  after the rename pad switched from a caret blink to a T9 pending-char.

Internal

• The on-device tests no longer guess which key they are talking to. Every
  tests/*.py script picked the first FIDO HID device the OS listed, so with a real
  YubiKey attached next to a board built VIDPID=Yubikey5 (both 1050:0407), tests
  10 and 15 ran against the YubiKey and reported its aaguid, its alwaysUv and
  its 6700 as RS-Key failures. Selection now lives in one place, tests/_device.py:
  the RSK marker breaks a tie, RSK_TEST_SERIAL / RSK_TEST_PATH name a target
  explicitly, and an unresolved choice stops the run instead of picking one. The
  seven copied find() helpers and ctaphid.find route through it, as do the
  python-fido2 suites (61, 65) and replug.reset_fido2 — that last one sent
  authenticatorReset to whatever it found first. Same bug class as the audit run-31
  fix in tools/rsk (ctaphid.find_all, connect_fido(exclusive=True)).
• The CCID half of that, over PC/SC. The reader pick was copy-pasted into 24
  scripts. Eighteen matched the RSK marker in the reader name but fell back to
  rs[0], so a build with USB_PRODUCT overridden drove whatever reader the OS listed
  first — next to a real YubiKey, the YubiKey, which enumerates ahead of the board on
  the maintainer's machine. 53 took rs[0] with no match at all, 90 took the first
  of the marker-matched, and 80_piv.py matched "Yubico" and "PIV" too, aiming a
  suite that blocks both PIN references and factory-RESETs the applet at a real
  YubiKey's PIV. All 24 now call _device.find_reader(), with RSK_TEST_READER as the
  pin. Five pass require_marker=True, so an unmarked reader reads as "not attached"
  rather than as the board: the destructive 80 and 90, and the reboot pollers 14,
  51 and 76, where "is the board back yet?" was answerable by a stranger — 51
  probes A0 00 00 05 27 47 11 17, Yubico's own management AID, and a real YubiKey
  answers it 9000.
• Test fixed. 31_openpgp_select.py asserted the OpenPGP VERSION (INS 0xF1)
  reply was 04 06 00 and so failed against any default build, which answers with
  the device firmware version (05 07 04). The expectation now derives from
  FW_VERSION like the firmware does, as does 10_fido_getinfo.py's
  firmwareVersion check, which had the default packed in by hand.
• Cross-executor Ordering consistency: the CANCEL_REQUESTED.store(false, …)
  false-clears in display/pin.rs and display/presence.rs are Relaxed (no
  publication occurs from a false store — the publication is the subsequent
  true/Release store), matching the same false-clears already in
  firmware/src/presence.rs.
• Test changed. t9_groups_are_printable_and_have_distinct_chars checked
  for duplicate characters within each group (where the old
  rename_charset_is_printable_and_cycles checked the whole charset). The test
  now also rejects a character appearing in two groups — a T9 char must belong
  to exactly one, else active_group/cycle_at on the rename screen is
  ambiguous. (const _: () = assert!(T9_GROUPS.len() == 10) pins the
  relationship hit_rename (Char(0..=9)) expects, so dropping a group fails
  the build instead of panicking on device at groups[gi].)
• The AA fringe in aa::filled_circle now blends to a caller-passed bg
  colour instead of the hardcoded theme::BG — truthful blending against
  the surface the circle is drawn over, so a future AA circle on a card or
  other non-BG region won't get a global-background halo. Existing call
  sites (boot splash, reset warning, PIN-pad dots, success circle) pass
  theme::BG, so the rendered output is byte-equivalent.
• firmware/build.rs strips a trailing # ... comment from a TOML value
  before parsing it (handling the "-quoted case, since a quoted value may
  legitimately contain #). Previously pin = 13 # GPIO for chip-select
  read as 13 # … and panicked parse_toml's u32 helper; today's four
  board configs are clean of inline comments, so this is a trap removed
  for future edits, not a current-data fix.
• firmware/build.rs renames the board-config display slice from b2 to
  disp_cfg and documents that display_cs is the semantic gate pin
  (a [display] section without cs is dropped, and the knobs fall back
  to the Waveshare defaults). It was previously a one-line clever and_then
  with no note explaining why.
• masked_entry's total reverts to the v0.4.4 one-liner
  (expected as usize).max(entered).min(ENTRY_MAX_SHOWN) — the UI redesign
  rewrote it as an if/else with the same result and a cosier comment
  ("no leftover outlines on delete") describing a change that didn't happen;
  the original is shorter and says exactly what it does.

Verify a download: https://github.com/TheMaxMur/RS-Key/blob/v0.4.5/docs/supply-chain.md