Andy Dixon
ec204f0ea2
Filesystem basics: mv now moves folders as well as files, and like mv(1) a NEW ending in / or naming an existing folder means into it. cp copies a file, or a folder with -r, by writing new records that share the original's objects, so it is instant and takes no space; rm already keeps objects another record still uses. New records go first, so an interrupted move leaves both names, never neither. get -r extracts a folder, keeping its structure. vault unlock [--for 5m] checks the passphrase and starts the binary again as a hidden __agent process, handing it the suite ID and seed over a pipe. The agent keeps them in mlocked memory, with core dumps off and (on Linux) PR_SET_DUMPABLE cleared, serves them on a Unix socket in a 0700 per-user directory whose owner is checked on every use, and wipes them and exits at a hard deadline or on vault lock. Commands ask it first unless VAULT_PASSPHRASE is set, and genkey stops any agent left for that path. Guide, man pages (new vault-cp, vault-unlock, vault-lock), README and screenshots updated; the guide's error table is split so it fits.
6.8 KiB
vault
A post-quantum encrypted file vault in a single binary.
Files are encrypted with a hybrid FrodoKEM-640 + ML-KEM-768 key encapsulation
and ChaCha20-Poly1305. The vault directory holds only randomly named
ciphertexts. Names, sizes, modes and timestamps live inside encrypted
catalogue records of one fixed size. File bodies are cut into 4 MiB objects,
the last padded to a power-of-two bucket, so every object is 64 KiB to 4 MiB
whatever the file's size. Object names derive from a secret seed in the
record, so nothing on disk links a file's pieces. Records and objects are
spread over 256 subfolders (records/3f/3fa9…) to keep directories small.
Build
go build . # produces ./vault
go test ./...
./build.sh # every OS and CPU architecture, into ./dist
build.sh cross-compiles static binaries for every target in
go tool dist list except iOS and 32-bit or x86 Android, which need cgo. It builds 32-bit ARM Linux for
both armv6 and armv7, and writes dist/SHA256SUMS. Set TARGETS to build a
subset, for example TARGETS="linux/amd64 darwin/arm64" ./build.sh.
User guide
A full illustrated guide lives in docs/guide. Build the PDF with
docs/guide/build-pdf.sh, which fetches Tectonic if it isn't installed.
docs/guide/shots.py regenerates its screenshots from the real binary.
Man pages
docs/*.1 holds vault(1) and one page per command (vault-add(1),
vault-mirror(1) and so on). Install them with
cp docs/*.1 ~/.local/share/man/man1/.
Use
vault genkey # create ~/.vault/key, protected by a passphrase
vault init # create ~/.vault/store
vault init --filler 2G # the same, with at least 2 GiB of filler
vault add report.pdf # stored as report.pdf
vault add report.pdf work/2026 # stored as work/2026/report.pdf
vault add ./photos # every regular file under photos/
vault add https://example.com/a.pdf # download and store a URL
vault migrate ./photos # like add, then delete the originals
vault mirror https://example.com/ # crawl a site into example.com/, links rewritten
tar -cz photos | vault absorb backup # every file in a tar (or .tar.gz) under backup/
vault absorb backup < photos.tar.gz # same, from an existing archive
vault ls # the top level: folders (ending in /) and files
vault ls work # one level down, inside work/
vault list # every file, all the way down, full names
vault list work # everything under work/
vault get work/2026/report.pdf # writes ./report.pdf
vault get report.pdf -o copy.pdf
vault get -r work # a whole folder, into ./work
vault mv report.pdf old/report.pdf
vault mv report.pdf old/ # the same: into a folder, like mv(1)
vault mv work archive/work # folders move too
vault cp -r work work-backup # instant; copies share their encrypted data
vault rm old/report.pdf
vault rm -r old # a whole folder; lists it and asks first
vault rm -rf old # the same, without asking
vault unlock # remember the key for 5 minutes (--for 30m)
vault lock # forget it now
Commands that change the vault take an exclusive lock on DIR/.lock, and
commands that read it take a shared one. A second process waits, and says
so, until the first finishes. The operating system releases the lock if a
process dies. Plan 9 and WebAssembly have no file locks, so there the lock
file itself is the lock. After a crash on those platforms you remove it by
hand; the error message names it.
Adding a name that already exists replaces it. get refuses to overwrite an
existing file and writes nothing until the whole object has authenticated.
--key PATH and --vault DIR work with every command. They default to
$VAULT_KEY and $VAULT_DIR, then to ~/.vault/key and ~/.vault/store.
The passphrase is read from the terminal, or from $VAULT_PASSPHRASE for
scripts.
vault unlock asks once, then starts a helper process, like ssh-agent,
that holds the unlocked key in memory until a hard deadline (5 minutes by
default) or vault lock. Until then commands skip the prompt, unless
$VAULT_PASSPHRASE is set, which always wins. The key reaches the helper
over a pipe and never touches the disk; the helper's memory is locked
against swapping, it cannot be core-dumped (or, on Linux, debugged), and it
answers only on a socket in a directory private to your user. While it is
unlocked, anything running as you can use the vault, decoy or not, so lock
before handing the machine over.
Decoy passphrase
vault genkey offers an optional second passphrase. In scripts, set
$VAULT_DECOY_PASSPHRASE. Each passphrase unlocks its own key, so a command
run with the decoy passphrase lists, adds and retrieves only decoy files. The
real files never appear.
vault genkey # set a passphrase, then a decoy passphrase
vault add plans.pdf # with the real passphrase
vault add shopping.txt # with the decoy passphrase
vault list # shows only what that passphrase stored
Two measures keep the real files deniable:
- The key file always holds two slots of identical size in random order. An unused slot is random bytes, so the file does not show whether a decoy exists or which slot is real.
initfills every vault with 8 to 32 random entries that nobody can decrypt. Anyone holding only the decoy passphrase sees unreadable entries in every vault, so hidden real files look the same as filler.
A decoy can only be set when the key is created. Someone who watches the vault over time can still see it grow when real files are added.
Default filler is only a few tens of megabytes, so someone holding the decoy
passphrase can see when far more data than that is unopenable. To hide more,
create the vault with vault init --filler SIZE (500M, 2G, 1T), at
least as large as what you will hide. Filler entries are written by the
same code as real files, with a spread of small and multi-chunk sizes, so
they look like a real collection. Filler can only be added at init;
added later it would stand out as a burst of new entries.
Limits
- The storage host can see how many records and objects exist, so roughly how much is stored, and when the vault changes. Chunking hides a file's size in any single snapshot, but objects written together can be grouped by anyone watching the folder or a sync service's upload log.
- Padding still costs space: small files round up to a 64 KiB-or-larger bucket and a big file's last piece to its bucket, at most 4 MiB per file.
- The key file is the only way in. Losing it or its passphrase loses the data.
Licence
GPL-3.0-or-later. The cryptographic core is derived from Fury.