Shizuku-routed key attestation calls reach keystore2 with callingUid
set to shell (2000) or root (0) instead of the originating app's uid.
target.txt has no entry for those uids so shouldSkipUid returned true
in onPreTransact, and handleGenerateKey was never entered. The
transaction reached the real KeyMint HAL, which on older Keymaster 4.x
HALs rejects Tag::ATTEST_KEY with -49 UNSUPPORTED_TAG.
Move the shouldSkipUid gate from onPreTransact into handleGenerateKey
itself, evaluated after attestationKey and isAttestKeyRequest are
parsed. Skip only when the request is neither BYO nor attest-key-
purpose. CREATE_OPERATION keeps its outer-level gate (not part of the
BYO flow).
After this change, Shizuku-routed BYO requests enter dispatch, hit
forceGenerate=true via the prior simplification, and route to
doSoftwareKeyGen. Non-BYO non-attest calls from shell/root uids
still fall through to HAL unchanged.
D3 (option 2B) from /home/rootdev/.claude/plans/breezy-seeking-wozniak.md.
Any attest-key request or BYO request goes software unconditionally.
Drops the alias-Elvis lookup and the nspace fallback added by the
d7dc5e0 -> 5f72acb -> 0b2c34f revert/restore churn, both of which
silently missed when callers (Shizuku, post-restart sessions) did
not share an in-memory KeyIdentifier(uid, alias) with the attest
key originally registered in attestationKeys.
The (shouldPatch && isAttestKeyRequest) clause is subsumed by the
broader isAttestKeyRequest clause.
D2 from /home/rootdev/.claude/plans/breezy-seeking-wozniak.md.
The race added in 8fdc59a forwarded BYO attest-key requests to real
HAL on cache miss, producing -49 UNSUPPORTED_TAG on devices whose
persistent attest key alias survived in keystore2 across daemon
restarts but never re-entered our in-memory attestationKeys set.
AUTO resolution now relies solely on
ConfigurationManager.getPackageModeForUid (config/ConfigurationManager.kt:115),
which uses DeviceAttestationService.isTeeFunctional
(attestation/DeviceAttestationService.kt:67) — a Kotlin by-lazy probe
evaluated once per daemon session. Matches upstream JingMatrix and
the v5.0-138 baseline.
Drops the AtomicReference<Boolean?> identity-equality compiler
warnings the race relied on.
After PR #22 and the AUTO-mode extension started caching attested
generateKey responses in teeResponses, KEY_ID getKeyEntry lookups for
attested keys returned from memory in ~1ms while non-attested keys
forwarded to real keystore2 took ~1.5ms.
TimingSideChannelProbe measured the 1.55x ratio against its 1.1x
threshold and flagged the asymmetry.
Forward non-attested generateKey to real keystore2 with post-hook
enabled (Continue instead of ContinueAndSkipPost), and extend the
GENERATE_KEY post-hook to cache no-chain responses into teeResponses.
The KEY_ID lookup added in the previous commit now resolves both
paths from memory at matched latency. Cert-chain patching is skipped
for the no-chain branch because there is no attestation extension to
rewrite.
PR #22's KEY_ID lookup at Keystore2Interceptor.onPreTransact only
scanned generatedKeys, which is populated exclusively by
doSoftwareKeyGen (GENERATE mode and the attest-key override path).
For AUTO packages on TEE-good devices the real TEE handles
generateKey and the response lands in teeResponses via the existing
GENERATE_KEY_TRANSACTION post-hook, so getKeyEntry(KEY_ID) by
TimingSideChannelProbe missed and the call leaked SSE.
Add findTeeResponseByKeyId companion helper that mirrors
findGeneratedKeyByKeyId's shape but scans teeResponses keyed by
response.metadata.key.nspace. Wire it as a fallback after the
existing PR #22 lookup. Behavior unchanged for GENERATE packages.
getPackageModeForUid collapses AUTO -> PATCH/GENERATE at the call site
based on DeviceAttestationService.isTeeFunctional, so isAutoMode never
observed Mode.AUTO and always returned false. That made the
isAuto-gated dispatch arms in KeyMintSecurityLevelInterceptor and the
entire raceTeePatch path unreachable.
Iterate packageModes directly with the same priority order as
getPackageModeForUid: first non-null mode wins. AUTO returns true,
PATCH and GENERATE return false. Activates raceTeePatch for AUTO
packages on the first generateKey per security level.
Real keystore2 SSE replies passed through SkipTransaction kept
the daemon's anyhow chain in the parcel string. Run those replies
through the same synthesizer used for module-generated SSEs so
wire shape stays consistent regardless of source.
StrongBox was pinned to attest/keymaster v300 regardless of SDK,
which mismatches devices shipping KeyMint v400 on Android 16.
Fall through to the same SDK_INT->version map used by the TEE
path so StrongBox reports the device-correct tier.
writeString(null) on service-specific exception replies left the
message word as 0xFFFFFFFF, which diverges from AOSP keystore2's
anyhow-formatted "Error::Rc(NAME)" / "Error::Km(NAME)" strings.
Map known ResponseCode/KeyMint codes to their canonical names so
the wire shape matches a stock TEE reply.
Reverts 5f72acb. The nspace attestation key lookup was part of
the score-4 baseline and closes a real NPE on KEY_ID-domain
references where the alias field is null. Reverted in error
during the score-4-to-14 rollback; restoring to match the
working baseline.
Reverts 22d1972. The updateAad SSE injection landed earlier as
the F1 quirk fix and was part of the score-4 baseline on mt6768.
It was reverted in error during the score-4-to-14 rollback; the
actual regression driver was the uncommitted DELETE_KEY absorber
which has already been dropped. Restoring to match the working
baseline.
Reverts 59836e1. Unconditionally injecting SSE(INVALID_TAG) on
non-AEAD updateAad matched the AOSP TA spec but diverged from
real-device behavior on mt6768, which returns silently. A
behavior-fingerprint detector on the Xiaomi probe flagged the
divergence and the Tamper score climbed from 4 to 14, with a
second detector raising key-tamper. Roll back to investigate a
device-conformant approach.
Reverts d7dc5e0. The KEY_ID-domain alias-null branch targeted
duck-detector's timing-side-channel WARN, but the WARN persisted
in subsequent testing and the combined fix attempts pushed the
Tamper score from 4 to 14 with a new key-tamper detection on a
second detector. Roll back to the a7e7e45 baseline to investigate
from a clean state.
The probe in duck-detector's TimingSideChannelProbe chain calls
generateSigningKey with a KEY_ID-domain attestation key reference
where alias is null. KMSLI.kt:498 fed that null alias into the
non-null String param of KeyIdentifier, triggering an NPE that
the outer runCatching wraps into a ServiceSpecificException(-49).
Any SSE crossing the binder boundary carries the Parcel.read/
createException(OrNull) stack frames, which the detector's
TeeReportReducer at line 2635-2641 matches verbatim to emit
"Captured private binder exception during timing skip".
Branch on the alias before constructing KeyIdentifier: when
alias is present, retain the existing isAttestationKey lookup;
when null (Domain::KEY_ID), scan attestationKeys for a matching
uid + nspace via generatedKeys. Mirrors AOSP keystore2's own
dispatch in database.rs (Domain::APP by alias, Domain::KEY_ID
by key id).
Real MediaTek mt6768 KeyMint silently returns OK on non-AEAD
updateAad, contradicting AOSP's mandate at
system/keymint/ta/src/operation.rs:430-446 to throw InvalidTag
when aad_allowed is false. Duck-detector flags this divergence
as "updateAad mismatch" in OperationErrorPathProbe.
Wire UPDATE_AAD into OperationInterceptor and inject
ServiceSpecificException(INVALID_TAG) when create params
indicate non-AEAD. AEAD (BlockMode.GCM) passes through to real
KeyMint untouched so AES-GCM round-trips remain valid.
Attested keystore operations finishing faster than non-attested
ones on the same device is a timing inversion that detectors
score against TEE coherence. Floor TEE op latency at 4ms to
preserve the natural ordering.
invalidate_on_error, avb_version, hash_alg, and size are sibling
props of vbmeta.device_state. When device_state is present but
its complements are missing, the partial set is itself a
detection signal. Fill with safe defaults if absent; never
overwrite when present.
Creating a vendor-specific boot prop on a device that never had
one is itself a detection signal. Existence-guard the four
boot-lock targets so absent props stay absent.
Live install on the user's Pixel surfaced a fourth bootloader-lock
prop the source plan never enumerated. The Chunqiu-style detector
card listed three indicators on its bootloader-unlock row and
flagged red because ro.boot.vbmeta.device_state remained unlocked
even after ro.boot.verifiedbootstate, ro.boot.flash.locked, and
ro.boot.veritymode were all spoofed.
Add ro.boot.vbmeta.device_state=locked to BootStateManager.targets.
Verified post-reboot on device: all four props now report the
spoofed values.
The source plan's bulletin-history schema declared a four-value
status enum: success, network_error, parse_error,
validation_rejected. The poller only ever wrote the first three;
when PatchLevelManager.updateTo silently rejected a date for bad
format, floor violation, past/future bounds, or atomicWrite IO
error, the history still recorded status=success and applied=true
because applied was set from isNewer before updateTo ran.
Make updateTo return Boolean. Wire the result through fetchAndParse
so a rejected apply lands as status=validation_rejected with
applied=false and an error string identifying the date that failed.
Closes the only spec gap from fancy-humming-firefly.md uncovered
during the source-plan cross-audit.
currentPatch returned the raw system= value unchanged. A
malformed value such as system=tomorrow flowed into the
lexicographic comparison `date > current`, where any well-formed
YYYY-MM-DD from the bulletin sorts before lowercase letters, so
the poller permanently judged its date "not newer" and never
applied an update. Validate the read value against the date
pattern; on mismatch, log a warning and treat as passive.
The 60-day window covers Pixel monthly bulletin cadence plus
pre-announcement slip but rejects far-future hostile inputs. The
prior bare constant left readers wondering whether the value was
arbitrary; the KDoc closes that loop.
applyToProps reaches Runtime.exec("resetprop", name, value) twice
per call, once for system and once for vendor. Boot-time
initialize, BulletinPoller's handler thread, and the PifObserver
inotify thread can all reach applyToProps independently. Two
concurrent invocations for different dates could interleave such
that system and vendor end up with mismatched values. Synchronize
on the singleton so each apply runs to completion before the
next begins.
A read failure (partial-write race during user edit, SELinux
denial, or any other IOException) used to fall through the
runCatching and rewrite the file with only the global block,
silently destroying every existing [pkg] override. Drop the
runCatching so the failure bubbles to atomicWrite, where the M3
guard in updateTo logs and returns without applyToProps, leaving
both file and props untouched.
isGlobalKeyAssignment treated any bare line whose first token
matched system/boot/vendor/all as a global assignment and
stripped it. A user line of literally "all" or "system" written
without a value (malformed config but reachable) thus got eaten
on the next atomicWrite. Require '=' in the trimmed line before
treating it as a key=value assignment.
parsePatchLevelValue silently synthesized day=01 when input was
6 chars (YYYY-MM) and caller wanted 8-digit YYYYMMDD. If
ro.vendor.build.security_patch ever returns YYYY-MM on a target
device (older Samsung firmware does), the synthesized day
disagrees with the real bulletin day -- a detection fingerprint.
Return null so getRealDevicePatchLevelInt falls through to
Build.VERSION.SECURITY_PATCH, which is always YYYY-MM-DD.
Revert 29b2a85. AOSP 15 KeyMint reference TA at
system/keymint/common/src/tag/info.rs:61-89 lists both Tag::EcCurve
and Tag::KeySize in KEYMINT_ENFORCED_CHARACTERISTICS, and
check_ec_params at common/src/tag.rs:632 says "Key size is not
needed, but if present should match the curve" -- the TA passes
through whatever the caller supplies and keystore2 supplies both
for EC keys per KeyMintBenchmark.cpp:234,259. Omitting KEY_SIZE
made the simulator's characteristics list shorter than real
hardware, a detection fingerprint.
PatchLevelManager.initialize ran once at boot; PIF edits required
a reboot to take effect. Add a FileObserver on
/data/adb/modules/playintegrityfix for CLOSE_WRITE, MOVED_TO, and
DELETE on the four known PIF filenames, re-resolving and applying
the new date when any of them changes. Skip the watch when the
PIF dir is absent so the daemon does not start a stale inotify
node before the module is even installed.
resolvePifPatch selected the last existing file regardless of
size. A 0-byte file picked up by lastOrNull caused JSONObject("")
to throw, the catch silently fell back to SystemProperties, and a
preceding non-empty PIF was ignored. Filter out zero-length files
so the lookup walks past them to the next valid candidate.
writeText and Files.move can throw IOException, SecurityException,
or AtomicMoveNotSupportedException. The exception previously
propagated through updateTo into BulletinPoller.fetchAndParse's
broad catch, which mislabelled it as "network_error" in the
history. Wrap atomicWrite, log the real failure, and return
before resetprop so the on-disk file and live props stay
consistent on failure.
atomicWrite previously overwrote the entire security_patch.txt
with only the three global lines, destroying the per-package
[pkg] overrides supported by ConfigurationManager. Read the
existing file, strip only global system/boot/vendor/all key
assignments, prepend the refreshed global block, and append
everything else (comments, blanks, all [pkg] sections) verbatim.
PatchLevelManager.initialize previously called updateTo, which
overwrote security_patch.txt with explicit dates and destroyed
the Phase 1 default of system=prop. Split prop application into
a new private applyToProps so initialize only resetprops; never
writes the file. BulletinPoller now treats currentPatch() == null
(the signal for system=prop or missing/blank) as passive and
skips updateTo. The file becomes user-owned config; props track
PIF or the device default.
BootStateManager.apply and PatchLevelManager.initialize ran after
initializeInterceptors, so keystore2 cached ro.boot.* and
ro.build.version.security_patch from the un-spoofed values during
hook init. Move both before the interceptor so the hook sees the
spoofed snapshot. ConfigurationManager stays between them since
it only loads files and is independent of prop state.
BulletinPoller.start ran inside App.main's outer try{...} catch
that rethrows, so any HandlerThread or Looper init failure killed
the daemon including keystore interception. Wrap the start call in
its own try so a poller failure logs and falls through, leaving
the rest of the pipeline alive.
fetchAndParse and appendHistory each catch their own exceptions,
but scheduleNext can throw IllegalStateException if the Looper is
torn down or any helper raises an unanticipated error. Without an
outer catch, the reschedule chain broke and the poller stayed dead
until reboot. Wrap the entire body so a thrown exception still
attempts to schedule the next poll.
PatchLevelManager only rejected dates more than ~1 year in the
past. A MITM serving <td>2099-12-31</td> from a spoofed bulletin
response slipped through validation and got written to
security_patch.txt plus resetprop'd. Add a 60-day upper bound past
today using LocalDate.plusDays so month boundaries are handled
correctly. The existing past bound stays.
BulletinPoller fetches the Pixel security bulletin index page on
its own HandlerThread with 5s/30s/2m/10m/30m bootstrap backoff,
then 24h steady cadence. The first <td>YYYY-MM-DD</td> match is
the latest published patch; newer-than-current dates flow through
PatchLevelManager.updateTo for validation + atomic write + resetprop.
Persists the last 10 attempts to last_bulletin_fetch.json (atomic
rename) with status, http_code, parsed_date, applied, and error
fields so operators can audit history without logcat.
Sepolicy rule appends TCP-socket allow rules for both ksu and
magisk source domains so HttpsURLConnection survives SELinux
enforcement on either root provider. Uninstall.sh cleans the
three new artifacts.
PatchLevelManager resolves the active security patch from
PlayIntegrityFix via the same six-path override chain as
Tricky-Addon's get_extra.sh (pif.json/pif.prop/custom.pif.*,
later entries override earlier ones). Falls back to live
ro.build.version.security_patch when no PIF source is present.
updateTo() validates YYYY-MM-DD format, rejects dates below
2020-01-01 or more than one year older than today, then atomically
stages security_patch.txt with explicit system/boot/vendor dates
and resetprops ro.build.version.security_patch plus
ro.vendor.build.security_patch. Cert tags 706/718/719 then encode
consistent dates via AndroidDeviceUtils.parsePatchLevelValue
(YYYYMM for OS, YYYYMMDD for VENDOR/BOOT per AOSP Tag.aidl).
Wired from App.main after BootStateManager.apply().
BootStateManager.apply() runs from App.main after ConfigurationManager
init and sets ro.boot.verifiedbootstate=green, ro.boot.flash.locked=1,
ro.boot.veritymode=enforcing via resetprop so the attestation
extension's hardcoded verifiedBootState=Verified agrees with what
detectors observe via getprop.
Adds an internal AndroidDeviceUtils.setProperty(name, value: String)
overload so the existing private ByteArray variant stays exclusive
to vbmeta digest persistence while config-package callers can set
plain string props without hex encoding.
Closes documented vulnerability D44 (countermeasure-matrix.md).
Previously module/update.json had to be hand-bumped to keep
versionCode and zipUrl in lockstep with module.prop's expanded
$gitCommitCount. Wire a refreshUpdateJson task to the
prepareModuleFiles${variant} pipeline so every zipDebug/zipRelease
regenerates the file from current verName and gitCommitCount.
Gradle's exec environment does not inherit the user's interactive
shell PATH, so cargo-ndk could not find cargo even when it lived in
~/.cargo/bin. Prepend ~/.cargo/bin to PATH for buildRustCertgen so
the Rust toolchain resolves reliably from any shell.
Java sourceCompatibility/targetCompatibility were already 21, but
the Kotlin compiler defaulted to JVM 17 bytecode, producing a
toolchain skew warning on every build. Align the Kotlin target to
match the Java target.
AOSP keystore2 attestation lists KEY_SIZE only when there is no
authoritative key-shape tag. For EC keys the curve already pins the
key size, so emitting both KEY_SIZE and EC_CURVE is a forgery
fingerprint. Guard the createAuth call accordingly.
Regenerated keys were being filtered as deleted because the
deletion marker in Keystore2Interceptor.deletedSoftwareKeys
survived past the regen call. Clear the marker at both software
and TEE generation paths so the next getKeyEntry returns the
fresh key instead of NOT_FOUND.
Instrumentation-only. Logs entry (primitive class, callingUid, input size)
and throwable propagation (class, SSE error code, message, stack top) in
both SoftwareOperation.updateAad and SoftwareOperationBinder.updateAad.
Intended to distinguish F1 hypotheses H1 (binder swallows SSE) vs H3
(AIDL signature drift) once duck's probe key is routed through our
simulator instead of the real TEE.
Address Copilot/CodeRabbit review feedback on the persistence PR.
1. SoftwareOperation: replace requireNotNull(keyPair) in SIGN/VERIFY/AGREE_KEY
branches with ServiceSpecificException(invalidArgument). The original
requireNotNull throws IllegalArgumentException, which the binder layer
wraps as KEYMINT_UNKNOWN_ERROR — defeating the goal of surfacing a
clean keystore-style error. Aligns with how ENCRYPT/DECRYPT already
handle missing key material in the same when block.
2. loadPersistedKeys: when a symmetric record has empty metadataBytes (e.g.
a save where Parcel.marshall() was empty for any reason), rebuild a
minimal KeyMetadata from PersistedKeyData primitive fields instead of
skipping the record. Skipping silently dropped the AES key, which is
the same 'logged out after reboot' behavior the PR is trying to fix.
The rebuilt metadata is structurally minimal but preserves the secret
material, which is the dominant correctness concern.
3. Comment fix: rebuildResponseFromRecord docs referred to 'v2 metadata
snapshot' — the format in this PR is v3.
Five issues that together caused keystore-pinned apps to be silently
logged out across reboots and config changes. All flow from the same
root cause: GeneratedKeyPersistence loses information on save -> reload.
1. Symmetric keys (AES, HMAC, 3DES) were never persisted at all
- GeneratedKeyPersistence.save only accepted KeyPair, ignoring SecretKey
- AndroidX security MasterKey (AES-GCM-256) regenerated on every
reboot, making EncryptedSharedPreferences undecryptable
- Apps that wrap session tokens in EncryptedSharedPreferences
interpret this as session expiry and force a relogin
2. Restored KeyMetadata authorizations differed from generation-time bytes
- loadPersistedKeys rebuilt KeyMintAttestation with mostly null/empty
fields, so toAuthorizations emitted a different tag set after
reboot vs. at generateKey time
- Apps that fingerprint metadata across keystore calls saw a
"changed key"
3. certificate / certificateChain split could shift after restore
- buildKeyEntryResponse called updateCertificateChain on the rebuilt
metadata, which is allowed to repartition leaf vs. chain bytes
- Apps with strict leaf fingerprint checks saw a "changed cert"
4. Touching ANY .xml under /data/adb/tricky_store wiped every cached key
- ConfigObserver called clearAllGeneratedKeys() which also calls
GeneratedKeyPersistence.deleteAll()
- Editing keybox.xml (or any unrelated .xml) thus deleted every
persisted key on disk
- Even the keybox-cache argument does not justify wiping per-app keys:
patched chains alone are stale, raw keypairs are not
5. SoftwareOperation NPE when restored keyParams missed PURPOSE tag
- Init dereferenced keyPair!! before checking purpose, so a
half-restored record crashed instead of producing a clean error
Single on-disk format (FORMAT_VERSION = 3) covers everything: PKCS8
private key bytes for asymmetric, raw secret bytes for symmetric, plus
the byte-identical KeyMetadata parcel snapshot so authorizations
restore exactly. Earlier dev-only formats are silently skipped by the
loader; the next generateKey for those aliases re-creates them in v3.
ConfigObserver now calls invalidatePatchedChains() instead of
clearAllGeneratedKeys() on .xml edits - only the chain cache is
stale, not the underlying keypairs.
Tested on OnePlus 13 (Android 16, KSU 3.2.4):
- Apps survive force-stop + cold reboot without losing keystore state
- Apps survive keybox.xml edits / replacements (touch, sed, cp -mv)
- Tamper score still 4 (CONSISTENT) on Duck Detector
- KeyAttestation chain output unchanged
Duck Detector's 'TEE Simulator generate-mode fingerprint' probe scans the
generateKey reply parcel for a 16-byte marker where the securityLevel byte
is 0x00 (SOFTWARE). Real KeyMint HAL uses 0x64 (KEYSTORE=100) for
keystore-enforced metadata (creation time, user ID, etc.).
This single-line change aligns with real hardware behavior and defeats
the probe. Tested on OnePlus 13 (Android 16, KSU 3.2.4):
- Before: 'TEE Simulator generate-mode fingerprint: Matched' (score 50)
- After: 'No TEE Simulator generate-mode fingerprint observed' (score 4)
Reference: https://github.com/eltavine/Duck-Detector-Refactoring/commit/e368038
v6.0 changed bare target.txt entries from AUTO to GENERATE, breaking
apps like BHIM that need TEE-backed attestation keys. Restore AUTO as
default and resolve it at config level (PATCH if TEE works, GENERATE
if not) to bypass the non-deterministic raceTeePatch path.
Enginex0