The wire format is a release candidate, frozen since 0.9.0 and not 1.0; the
current release line is 0.10.x. The implementation has never been audited by a
third party, and the hybrid KEM follows an IETF draft that is not yet an RFC.
Treat every container as a test artifact. The CLI requires --experimental
for this reason.
Only the latest release line (0.10.x) receives security fixes. Nothing is backported to earlier minors.
HIDE implements no primitives of its own and defines no novel KEM combiner. It
composes crates from the RustCrypto and hpke projects, all pinned to exact
versions:
- Key encapsulation: HPKE base mode with X-Wing (X25519 + ML-KEM-768, FIPS 203), HKDF-SHA256, ChaCha20-Poly1305.
- Payload: ChaCha20-Poly1305 in 64 KiB authenticated chunks; HMAC-SHA256 over the header.
- Signatures: hybrid Ed25519 + ML-DSA-65 (FIPS 204); both halves must verify.
- Key files: Argon2id with an enforced parameter floor and ceiling (memory 8–256 MiB, at most 4 lanes, 1–64 passes), then ChaCha20-Poly1305.
- Transparency: RFC 6962 Merkle tree, inclusion and consistency proofs.
- Group messaging: MLS (RFC 9420) through
mls-rs, classical X25519 suite only — post-quantum MLS suites are still an IETF draft.
unsafe_code = "forbid" applies to every crate except hide-ffi. A flaw in an
upstream primitive belongs upstream, though a report here is still appreciated
so the dependency can be pinned or replaced.
The crates that carry the post-quantum and HPKE code are not audited either.
The READMEs of ml-kem, ml-dsa and x-wing each say the implementation "has
never been independently audited"; the hpke README says that "nobody has
performed a paid audit of this crate" and records one internal security review,
by Cloudflare, of version 0.8; the mls-rs README says it "has not yet received
a full security audit by a 3rd party". Pinning a version fixes which unaudited
code runs; it does not make it audited.
HIDE defends a container against anyone who is not a recipient: they cannot read
it, alter it, truncate it, reorder its chunks, or learn its filename. A signed
container additionally proves which key produced exactly this content, even to
another recipient. It does not bind a key to a person, hide who or how many
the recipients are, hide the file size, or protect a key that an attacker can
read from process memory. The full model, including what a malicious recipient
and a malicious log operator can do, is in
docs/threat-model.md.
Known and documented, not new findings:
- An identity is not a person. The device log proves which devices an
identity trusts over time; nothing proves the identity belongs to a given
human. There is no directory, and no key-transparency service — the log
primitive exists and witness cosignatures on checkpoints are verified
(
spec/hide-1.md§16.2), but no one operates a log or a witness. - Sender authentication only when signed. A recipient of an unsigned container holds the CEK and can rewrite the recipient set and recompute a valid header MAC. A signature closes this; its absence does not.
- Forward secrecy is by erasure, not by ratchet. It exists only if the holder actually destroys the epoch secret. The CLI keeps epoch secrets in a passphrase-sealed store and erasing reseals it without the epoch, but an older copy of that file (backup, snapshot, remapped flash blocks) still holds what it held.
- No recipient anonymity. Ciphertext size and recipient count are visible; metadata is encrypted, not hidden. A public signature reveals the signer.
- No hardware protection. Argon2id resists an attacker who copies a key
file, not one who reads memory while the key is in use. There is no Keychain,
TPM, Secure Enclave or Keystore integration, and
keygen --insecure-plaintextwrites an unencrypted key on request. - A forgotten passphrase is unrecoverable. There is no escrow or reset.
- Group messaging is not post-quantum, and
mls-rsis as unaudited as the rest. The credential binding described below was introduced in 0.7.0 and has been parsed by no implementation outside this repository. - SSH authentication is not post-quantum.
hide agentoffers the Ed25519 half only. OpenSSH 10.6 (2026-10-06) enables the hybrid signature algorithmssh-mldsa44-ed25519(https://www.openssh.org/txt/release-10.6), but HIDE signs with ML-DSA-65 and its SSH key is Ed25519 alone, so HIDE's SSH support stays classical. It avoids a plaintext key on disk; nothing more. - An agent is a signing oracle. Anything that can reach the endpoint can request a signature, which is why confirmation is the default.
- Replay is only detectable by the verifier. A replayed challenge answer is a genuine signature; a verifier that keeps no spent-nonce record gains nothing over a plain signature.
- Advisory HIDE-2026-001:
a 95-byte key file could make
openallocate up to 4 GiB (affected 0.2.0–0.7.x). Found by the nightly fuzzer after 0.7.0 shipped, which is why it has an advisory and the 0.7.0 findings below do not. Memory is now capped at 256 MiB and parallelism at 4, before any allocation. - Identity and epoch decoders no longer reserve memory from the declared entry
count; MLS messages and
hide unsealinput are capped at 1 MiB; au64length or index is refused on 32-bit targets instead of being truncated. - Sixteen fuzz targets cover every parser of untrusted bytes — the container, key files, identity logs, epoch chains, transparency proofs, MLS messages, the ssh-agent decoder, public-key armor, the MLS device credential, the relying-party verifiers, and every byte-taking entry point of the C ABI — each fuzzed four hours every night with the corpus carried forward.
Found after 0.8.0 shipped: the identity-log transport encoding declared six
fields per entry and wrote seven. Rust read its own output, so no HIDE surface
was affected, but any other CBOR decoder misread the log. It is an
interoperability defect, not a security one. 0.9.0 writes array(7) and still
reads the 0.6–0.8 form (spec/hide-1.md §10.3).
An internal review — not a third-party audit — found the following. All are fixed in 0.7.0, and each fix has a test that fails on the previous code. They are listed because their existence is the best evidence that others remain.
- MLS credentials were self-asserted. A group member could present another
device's id in its credential; nothing proved it held that device's key. The
credential is now a HIDE-signed binding over the MLS signature key
(
spec/hide-1.md§13), verified on every roster change. - 7 of 8 signature stanzas were never verified. The header admitted up to
eight public signatures but checked only the first, so a recipient could
append stanzas nobody examined.
MAX_SIGNATURESis now 1 and a second stanza is rejected, not skipped. - One metadata key and nonce sealed two plaintexts in the confidential-signature path (the metadata before and after the signature was inserted). No disclosure resulted, but nonce reuse under one key is a fragility that should not exist.
- The identity-log encoding was not canonical: two distinct event sequences could serialise to the same bytes. Every variable-length field is now length-prefixed.
panic = "abort"madeHIDE_ERR_PANICunreachable in the C ABI: a panic inside the library aborted the host process instead of returning the error code the header documents. The FFI now builds withpanic = "unwind"and catches at the boundary.HIDE_LIBRARYwas an unconditional library-injection vector in every SDK: any process able to set an environment variable could replace the cryptographic core. It is now honoured only whenHIDE_ALLOW_LIBRARY_OVERRIDE=1is also set.- The ssh-agent read its confirmation from stdin, not the terminal, so a
piped process could answer for the user; and on Windows the agent's named
pipe was created without an ACL. Confirmation now reads the controlling tty.
On Unix the socket lives in a directory created with mode 0700. On Windows
the named pipe is still not restricted to the current user
(
apps/hide-cli/src/transport.rs); instead--no-confirmis refused there, so every signature must be confirmed on the terminal. - Argon2id parameters have a floor. A key file declaring weaker parameters than the floor is refused rather than opened.
Report privately through either channel:
- GitHub private vulnerability reporting: https://github.com/hide-protocol/hide/security/advisories/new (the Report a vulnerability button under the Security tab). The form is access-controlled, so prefer it for anything with sensitive detail.
- Email: hide-security@codai.ro. No PGP or HIDE key is published for it, and mail is not end-to-end encrypted, so keep sensitive detail out of email or send only a pointer and continue in the GitHub advisory.
Do not open a public issue for a suspected vulnerability. Public issues are for questions that are not sensitive. The machine-readable contact file is https://hide-protocol.github.io/hide/.well-known/security.txt (RFC 9116).
Include the affected version or commit, a reproduction (a container plus the steps), and what you believe the impact is. A failing test is the most useful form.
HIDE has one maintainer. A report is acknowledged as soon as he reads it; there is no guaranteed response time. We aim to publish a fix or a mitigation within 90 days and to coordinate disclosure with you; that is an aim, not a guarantee. If you have had no reply at all after 14 days, treat the report as unread and send it again through the other channel. Only the latest release line is fixed (see "Supported versions" above).
No bug bounty exists. A finding is credited in the CHANGELOG and in
docs/advisories.md unless you prefer otherwise.
Good-faith security research against your own HIDE instances, keys and data is authorised, and we will not pursue or support legal action over it. Do not access, alter or exfiltrate data belonging to anyone else, do not degrade a service other people rely on, and give us a reasonable window to fix what you find before publishing. If you are unsure whether something is in scope, ask first through either channel above.
- A container that decrypts under a key it was not encrypted to.
- A tampered container that authenticates successfully.
- Any input that causes a panic, an unbounded allocation, or non-termination in
a parser — the fuzz targets under
fuzz/are the starting point. - Divergence between the Rust implementation and
conformance/node/verify.mjs, since either side may be the wrong one. - Secret material reaching a log, an error message, or a file on disk.
- Identity-log forgery: an entry accepted as authored by a device that was not trusted at that position in the log, or a revoked device regaining authority without the recovery key.
- Epoch-chain rewriting: a chain that verifies while omitting, reordering or replacing an epoch that was published.
- Consistency-proof forgery: a proof accepted between two checkpoints where the smaller log is not a prefix of the larger.
- MLS binding bypass: a roster entry accepted whose credential does not bind the MLS key to a device currently in the identity's membership.