SIGNET WALLET FOR TEZOS

· Audit & remediation review

Source code security audit

A review of Signet for Mac, Windows and GTK, covering application key handling, transaction intent, persistence, recovery and code quality.

Audit performed by OpenAI Codex.
Source review, offline reproductions and targeted fix verification.

Current position. Source fixes are present for all five application findings. The private-file ACL regression has now executed successfully on Windows. GTK has native build/test and dashboard/Settings smoke evidence on Linux. The Windows logging-test follow-up is now resolved on a test-only fix branch: 505 native full-suite tests passed with no failures; 14 remain explicitly skipped. The test-only fix is merged into main on both checkouts.

Status includes subsequent native validation performed by OpenAI Codex over SSH. The exact tests, skipped cases and remaining GUI/platform limits are recorded below.

01 / Application findings

Problems and resolution status

The table separates the reported defect from the evidence used to close it. “Source fix” means the repair is present, with verification limited to source review or portable checks. The final row tracks the separate Windows test-reliability follow-up.

Fixed Reported path repairedSource fix Platform coverage limitedValidation due Native checks still neededNot affected Original finding does not apply

Scroll each table horizontally to see all platforms.

Application findings and validation follow-up — latest recorded status.
IDProblemMacWindowsGTKResolution / remaining work
G03Unbounded faucet workNot affectedFixedBounds & cancellation testedNot affectedUnreasonable work is refused; cancellation and time budgets reach requests and hashing.
G06 / T03Token transfers refused by host policyNot affectedNot affectedFixedNative policy tests passedGTK now classifies token transfers as signing calls; its policy matches its own upstream pin.
G07Private permissions for custom wallet foldersNot affectedFixedNative Windows ACL test passedNot affectedProtected DACL creation/replacement is implemented, and the unchanged regression passed on Windows. Network-share providers were not exercised.
K02Unredacted crash-exception loggingNot affectedFixedFormatter & rotation testedNot affectedThe unhandled-exception path now redacts each line before writing and rotates the crash log.
T04Overlapping asynchronous signing callsGate presentFixedNative gate tests passed · 4327c79Gate presentA bounded FIFO gate holds each foreground call until its asynchronous outcome settles.
TestShared logging-test selector
Test-reliability follow-up
Not applicableFixedNative full run: 505 passed, 0 failedNot applicablePrivate bridge and temporary log committed as 09d580e. All original assertions retained; native focused and full parallel runs pass. Merged into main on both checkouts.

02 / Experimental feature

Experimental Shielded Transaction support

Experimental Shielded Transaction support is not for production use, and cannot be used on mainnet. Every item in this table concerns that experimental feature, including its separate spending keys, services, counters, proofs and backups.

Scroll horizontally to compare Mac, Windows and GTK.

Experimental Shielded Transaction support only — non-production; unavailable on mainnet.
IDProblemMacWindowsGTKResolution / remaining work
G01 / T02One-time signing-index collisionsFixedProcess & alias probes passFixedNative lock & alias tests passedFixedNative tests & portable probesCanonical process-local locks and file locks protect the complete counter reservation.
G02Plaintext shielded spending keysOpenOpenOpenSeparate shielded master keys remain unencrypted in wallet JSON and backup generations. Warnings were added.
G04Unencrypted shielded-service trafficOpenOpenOpenThe default experimental operator and rollup services still use HTTP. Warnings do not encrypt traffic.
G05Wallet files accepted by writers but refused by readersFixedFixedNative store/backup tests passedFixedNative backup tests passedThe store, snapshot and fingerprint readers now share a 128 MiB shielded-file limit.
G08Untrusted rollup numbers could trapFixedNot affectedFixedNative regressions passedProtocol-range checks and checked arithmetic repair the reported Swift paths. The original trapping expression did not apply to Windows.
G09New shielded state missed scheduled backupsImplementedImplementedPresent in pinBackup scheduling now follows essential key, address, pool and submission-record mutations.
T01Uncertain submissions lost their records or reservationsFixedFixedNative retry regressions passedFixedNative retry regressions passedPayloads are saved before submission. Exact-payload retries preserve uncertainty when a later attempt is refused.
RecoveryRestoring an older one-time signing counterPolicy pendingPolicy pendingPolicy pendingBackup copies include the counters. A safe restoration policy still needs definition and verification; no rollback exploit was demonstrated.

The counter-restoration row is a remaining policy and verification requirement, not a demonstrated rollback exploit. Backup scheduling being implemented does not guarantee a successful snapshot when storage fails.

03 / Evidence and scope

What was checked

Latest recorded checks; selected test counts are not full-suite coverage.
PlatformReviewed sourceResultsLimits
Macb6297c6Same complete source tree as retry-fix commit 555311f.56 tests in 12 native suites passed after the retry repair. Storage and logging regressions passed.Selected native suites. Live signing and full GUI acceptance were not performed.
Windows09d580eTest-only fix merged into main on Windows and the local checkout, based on 4327c79. Recorded and checked-out upstream: b6297c6. Native Windows x64 host.Full native parallel suite: 505 passed, 0 failed, 14 skipped out of 519. Focused logging/console checks: 7 passed. The earlier native ACL/gate regressions and unchanged production WinUI build passed.The clean full run applies to this same commit, now on main in both checkouts. Interactive Windows GUI and network-share providers remain untested.
GTK / Linux4b87d86Upstream: b6297c6. Debian 13, Linux x86_64, Swift 6.4.2.Native app/test build and both native test runs passed. Runner reports: 428 tests / 130 suites and 2 port tests / 1 suite; 42 explicit skip notices, no failures. Dashboard and Settings smoke checks passed on private Xvfb.Reported totals include skipped cases. No live transfers or hardware acceptance. FreeBSD was not run; optional reference fixtures were absent on this Linux host.

Windows follow-up closed at the verified revision: 09d580e isolates LoggingPolicyTests.ConsoleLinesFromJavaScriptAreRedactedAndCapped in its own bridge and temporary log, selects its single matching echo, and retains every redaction/length assertion. Native focused checks passed 7/7; the normal parallel full run passed 505 tests with no failures and 14 explicit skips. The earlier 504-pass/1-failure result remains historical evidence. The test-only commit is merged into main on both checkouts; no push was performed.

Test isolation: native Windows ACL/gate tests and the WinUI build used unchanged source. Only default profile/local-data path resolution was redirected in a temporary export for the full Windows suite, keeping real user state untouched. GTK used unchanged source with synthetic home/temp mounts and network isolation; the user's desktop was not addressed. Overlapping selected and full-suite test counts must not be added.

Experimental Shielded Transaction support verification: The earlier required bridge-vector check passed 173 cases on macOS, and its missing-fixture negative control failed as expected. The optional reference fixture was absent on the native Linux host, so no additional native vector coverage is claimed. Full authentication-root checks remain outstanding. All storage checks used synthetic state.

04 / Remaining work

What is still outstanding

AreaRequired workClassification
Windows GUI and provider coverageNative ACL/gate regressions and the WinUI build have passed. Interactive Windows GUI acceptance, real-device flows and network-share filesystem providers remain outside the completed checks.Further coverage
GTK further coverageLinux native build/tests and dashboard/Settings smoke checks have passed. FreeBSD, live operations and real devices were not exercised.Further coverage
GTK QR fixturesAdd independently generated expected QR modules and prove that a controlled encoder mutation is detected.Quality follow-up
Windows policy regressionKeep a permanent test comparing the Windows signing declaration with its own recorded upstream pin. The audit comparison matched 20 entries.Quality follow-up
Windows collectible refreshMeasure refresh work before deciding whether an optimization is justified. No application speedup has been demonstrated.Measurement due
Experimental Shielded Transaction supportAddress plaintext keys and HTTP service traffic, define safe counter restoration, enforce required native vector-section/schema completeness, and run full authentication-root checks.Experimental only

Code-quality follow-ups are not evidence of successful exploitation. The original code-quality review used eleven independent review lenses and a separate synthesis; later rechecks verified implementation status rather than rerunning those lenses. Complexity and complete mutation-coverage targets remain unmeasured. No broad rewrite was justified by file size alone.

05 / Supporting detail

Finding details and corrections

G03 · Windows · Fixed

Bound faucet work and propagate cancellation

The original solver accepted a difficulty of 65 even though a SHA-256 hexadecimal digest has only 64 characters. A request could therefore leave hashing work running indefinitely. Difficulty, challenge counts and lengths are now checked before work begins. Cancellation and solve/total budgets reach requests and hashing. Focused limit, cancellation and deadline regressions pass.

G06 / T03 · GTK · Source fix verified

Restore token-transfer classification

The GTK host omitted sendTokenTransfer from its signing-call set, so legitimate NFT transfer and burn actions were refused before execution. The entry is restored. The new port tests compare the declaration with its own pin and now pass natively on Linux; the independent inventory and removal control also passed. Live NFT transfer/burn was not exercised.

G07 · Windows · Native validation due

Create sensitive files with private permissions

The earlier writer relied on inherited directory permissions. The implementation now creates each sensitive temporary file with a protected DACL allowing the current user and SYSTEM before writing, then replaces the target. Public artwork opts out. The test that returned immediately on macOS has now executed its actual Windows assertions and passed. It checks protected wallet-file permissions under a permissive parent, replacement permissions and the artwork opt-out. The creation API supplies the DACL before writing; a separate first-byte observation and network-share check were not added.

K02 · Windows · Fixed

Redact unhandled exceptions before persistence

The old unhandled-exception handler appended complete exceptions directly to a crash log. It now uses CrashLog.Append, whose formatter redacts every exception line before writing and rotates the file. Synthetic nested-exception and rotation tests pass. The original finding was a hardening gap; a real spending-key disclosure through this path was not demonstrated.

T04 · Windows · Fixed in 4327c79

Hold the bridge turn until the asynchronous result settles

A dedicated engine thread serialized invocation but allowed two pending asynchronous signing calls to overlap. A benign test export confirmed maximum concurrency of two; it did not demonstrate transaction or key substitution. The correction adds one foreground call per bridge with up to 64 FIFO waiters. Timer, fetch and device callbacks continue on the engine thread while that call awaits its result.

Read deadlines begin only when admitted. Success, rejection and timeout release the turn; a timed-out read resets the context before the next call begins. Abandoning the caller’s wait does not release an unresolved signature. Regressions cover ordering, synchronous and asynchronous failure, timeout recovery, queue saturation and independent bridge instances. The change is merged into local main; no push was performed during this audit.

Validation follow-up · Windows · Fixed in 09d580e

Logging test now owns its bridge and log

LoggingPolicyTests.ConsoleLinesFromJavaScriptAreRedactedAndCapped previously read the last echo from TaquitoBridge.Shared, which another parallel test also used. The original full native run failed its replacement-marker assertion; the preceding assertion excluding the original synthetic key and hex string passed. No raw-key disclosure was demonstrated.

The test-only correction creates a private bridge and unique temporary log, then asserts that exactly one matching echo exists in that instance. All original key-exclusion, marker and line-length assertions are retained. No production logging code changed. The native focused run passed 7/7, followed by a full offline run with normal parallel execution: 505 passed, 0 failed, 14 skipped.

The fix, 09d580e564a96efea5ef53fb1b9a3c5783ea817b, is merged into main on both machines. The follow-up is closed at that verified revision; no push was performed. Historical failing results are preserved.

G01 / T02 · Experimental Shielded Transaction support · Reproduced cases repaired

Reserve each one-time signing index atomically

Process-local locks alone allowed separate writers to reserve the same WOTS+ index. A later path-alias probe also found two process-local lock identities for one file. Canonical lock identity plus a file lock now protects the complete update. Actual-source probes reserved indices 0 and 1 and persisted 2. Subsequent native Windows and Linux lock/alias regressions passed. Network-share providers and mixed-version writers remain outside the checked scope.

G02 · Experimental Shielded Transaction support · Open limitation

Separate shielded keys remain plaintext

The experimental feature generates its own master spending key. It remains directly serialized in wallet JSON and backups, independently of a Tezos account password. User-facing warnings now explain that distinction. File permissions reduce access but do not encrypt an obtained copy. Encryption remains an experimental-feature improvement, outside production/mainnet scope.

G04 · Experimental Shielded Transaction support · Open limitation

Experimental service connections still use HTTP

The shipped experimental operator and rollup services use HTTP. Traffic, including the operator bearer token, lacks transport confidentiality. Warnings identify these connections; they do not supply encryption. This observation concerns the services used by Experimental Shielded Transaction support, which cannot be used on mainnet. No live interception or token-privilege exploit was performed.

G05 · Experimental Shielded Transaction support · Reported mismatch repaired

Keep store and backup read limits consistent

The original writer could create state that its reader refused. An intermediate repair still left backups limited to 32 MiB while the store accepted 128 MiB. The store, backup snapshot and fingerprint paths now share the shielded-file bound, and excessive writes fail before replacement. Synthetic files of about 38 MB save and reload; complete backup regressions above 40 MiB pass in the native Mac, Windows and GTK/Linux checks.

G08 · Experimental Shielded Transaction support · Reported paths repaired

Validate untrusted levels before arithmetic

A Swift fee path added to an untrusted head level before clamping it, allowing an extreme integer to trap. Levels are now protocol-range checked and related fee additions report overflow. The reported paths and error cases pass the native regressions. The original exact expression did not apply to Windows, which widened its integer before addition.

G09 · Experimental Shielded Transaction support · Scheduling implemented

Schedule backups after essential mutations

The backup file list included shielded state, but a startup snapshot could predate a newly generated key or receive address. Backups are now scheduled after essential creation, reservation and pre-submission record paths. Windows also backs up after reserving a signing index. Storage failure can still prevent completion; this fix does not guarantee a fresh snapshot after every possible state transition.

T01 · Experimental Shielded Transaction support · Fixed

Preserve uncertainty across rejected retries

The exact payload and its record are now persisted before contacting the operator. Transport and transient polling failures keep note reservations, and the interface offers an exact-payload retry. A refusal of a later attempt cannot prove that the earlier attempt was never accepted, so retry refusals preserve the original record. The regression simulates acceptance with a lost reply, rejected retries for both operations, retained reservations and later replay after correcting the token.

Recovery · Experimental Shielded Transaction support · Policy and verification pending

Where the signing counters live

Each shielded address has its own next-unused WOTS+ index, serialized as accounts["<tz-account-address>"].addresses[i].nextKeyIndex in tzel-<chain>.json in the selected wallet directory. The default Shadownet paths are:

Mac / GTK: ~/.signet/tzel-shadownet.jsonWindows: %USERPROFILE%\.signet\tzel-shadownet.json

The increment is persisted before signing. The adjacent lock file coordinates access; authentication-tree binary files contain rebuildable leaves. Neither replaces the JSON as counter state. Backup generations include that JSON, so restoring an older generation can restore an older index. A restoration policy must preserve one-time-key safety. The concurrency fixes alone do not establish that policy, and no destructive restore or successful forgery was attempted. These observations concern Experimental Shielded Transaction support, which cannot be used on mainnet.

Audit attribution and method

OpenAI Codex performed this audit and the recorded follow-up verification. The work combined source inspection, exact upstream-pin comparisons, synthetic offline reproduction, selected Mac tests, subsequent native Windows and GTK/Linux builds/tests, private-display GTK smoke checks and review of fixes. Codex also implemented the targeted Mac retry correction and Windows asynchronous-call gate described above.

Initial source baselines were Mac 86b28df, Windows b853993 and GTK 671ffdf. The summary uses the later reviewed revisions and fixes, rather than treating historical failures as still open. Native tests used synthetic state; original source checkouts stayed unchanged during validation. No real wallet, existing user log, live credential, live device or transaction injection was used. This scoped source-and-test review does not establish exhaustive application safety; the platform and verification limits above remain part of the result.