kerbcore
Pure-Rust Kerberos building blocks — no FFI, no dependency on a host krb5,
#![forbid(unsafe_code)].
kerbcore implements the modern Kerberos encryption-type matrix, a from-scratch
DER / RFC 4120 message codec, and a plain AS / TGS client — the protocol core a
defender, DFIR engineer, or interoperability author needs without linking a system
krb5. The crate is pure: bytes in, bytes out. Network I/O stays with the caller.
Install
[]
= "0.2"
Requires Rust 1.88+.
Status
Pre-1.0 — 0.2.x is the current line (typed-key enctype dispatch, AP-REP mutual auth,
GSS RFC 4121, SPNEGO, FAST, KKDCP, kpasswd, cross-realm referrals). A 0.3.0
API-freeze candidate is planned before 1.0. The AS/TGS crypto, codec, and client are
complete and live-validated end-to-end against Windows Server 2019, 2022, and
2025 KDCs: each obtains a real TGT and service ticket, handling the three different
salt schemes the servers return (the salt is parsed from the KDC's ETYPE-INFO2,
never guessed).
When to reach for kerbcore (honest table)
Kerbcore covers a specific niche in the Rust Kerberos ecosystem. It is not the right choice for everyone; the alternatives are legitimately better for other use cases.
| You want... | Reach for | Why |
|---|---|---|
| Kerberos client that just works everywhere your OS runs | libgssapi-sys |
Battle-tested MIT/Heimdal FFI; largest user base; the boring safe default |
| Windows-native SSPI parity from Rust | sspi-rs |
Real SSPI reimplementation; substantial user base; handles Windows quirks |
| Just KRB message codec, no client | picky-krb |
Smallest surface; long history; used by sspi-rs internally |
| Static binary, no libclang at build, no libkrb5 at runtime | kerbcore | Zero FFI, zero system deps |
| Cross-compile to musl / minimal Alpine / embedded | kerbcore | Pure-Rust, #![forbid(unsafe_code)], tiny external tree (8 deps) |
| Wire-level control (roast, forge, inspect PA-data, walk KDC-REP by hand) | kerbcore | Types are yours; the codec is total (no panic on hostile input) |
| Environments where the MIT/Heimdal C audit surface is unacceptable | kerbcore | Pure-Rust, no unsafe |
| A production-mature Kerberos crate with a large ecosystem | not kerbcore yet | Pre-1.0, small user base — pick libgssapi or sspi-rs |
The pure-Rust-Kerberos-with-no-FFI category is a small niche. Kerbcore's goal is to be the honest top of that niche, not to displace libgssapi.
Examples
Three runnable examples under examples/:
get_tgt— obtain a real TGT from a live KDC (env-configured), from scratch in ~90 LOC. AS-REQ → KRB-ERROR (salt) → PA-ENC-TIMESTAMP → AS-REP → decrypt session key.asrep_roast— no-preauth account → hashcat-m 18200line. The offline-cracking primitive behind AS-REP roasting, in ~90 LOC and no runtime.gss_wrap_unwrap— RFC 4121 message protection round-trip (offline, no KDC needed). Shows the post-handshakeWrap/Unwrap/GetMIC/VerifyMICprimitives you'd otherwise reach for through libgssapi.
Used by
ccache-io— MIT ccache v4 reader/writer. Itskerbcorecargo feature bridgesPrincipal/KeyBlocktoPrincipalName/EncryptionKey, so callers who already speak kerbcore fold ccache I/O in with no manual field-shuffling.ms-pac-forge— MS-PAC forge / verify. Uses kerbcore as a differential oracle for the RC4-HMAC path (dev-dependency today; runtime dep after ms-pac-forge's typed-error refactor).adhammer— AD security assessment CLI. Consumes kerbcore's crypto + codec + client for AS-REP roast, TGS-REP kerberoast, S4U2Self / S4U2Proxy (RBCD), Golden / Silver / Diamond forge, and PKINIT.
Encryption-type coverage
The modern Active Directory encryption-type matrix — not a single profile:
| etype | profile | status |
|---|---|---|
| 17 | aes128-cts-hmac-sha1-96 (RFC 3962) | ✅ |
| 18 | aes256-cts-hmac-sha1-96 (RFC 3962) | ✅ |
| 19 | aes128-cts-hmac-sha256-128 (RFC 8009) | ✅ — RFC 8009 §7 KAT |
| 20 | aes256-cts-hmac-sha384-192 (RFC 8009) | ✅ — RFC 8009 §7 KAT |
| 23 | rc4-hmac (RFC 4757) | ✅ — differential-tested |
| 1 / 3 | des-cbc-* | out of scope — no weak/legacy crypto shipped |
RFC 8009 (etypes 19/20) is implemented from scratch — few Rust libraries do. On Windows these are added in Server 2025 / Windows 11 24H2 (disabled by default); MIT krb5, Heimdal, and Samba negotiate them today.
ASN.1 / RFC 4120 message codec
A hand-rolled DER (X.690) codec — no external ASN.1 crate — with a total,
no-panic decoder: every malformed byte returns a DerError, never a crash.
der— canonical DER encoder + total decoder (KAT'd integer/length encodings).types—PrincipalName,KerberosTime,EncryptionKey,EncryptedData,PaData,Checksum,Realm.messages—Ticket[APP 1],KDC-REQ(AS-REQ / TGS-REQ),KDC-REP(AS-REP / TGS-REP),KRB-ERROR[APP 30].
Every message round-trips, application/context tags are verified, and Ticket,
AS-REP, and KRB-ERROR encode byte-identically to picky-krb (kerbcore DER →
strict decode → re-encode → same bytes). picky-krb is used only as a dev-dependency
differential oracle.
Crypto primitives
The RFC 3961 / 3962 primitive set, key-length-generic across the AES-SHA1 profiles:
nfold— RFC 3961 §5.2 n-foldaes_cts_encrypt/aes_cts_decrypt— RFC 3962 §5 AES CBC-CTS (CS3), AES-128 and AES-256dr/dk— RFC 3961 §5.1 DR/DK derivationhmac_sha1_96— RFC 2104 HMAC-SHA1 truncated to 96 bitsderive_kc/derive_ke/derive_ki— RFC 3961 §5.3 subkey derivationencrypt_message/decrypt_message— the generic encrypt-then-integrity primitivestring_to_key— RFC 3962 §4 (PBKDF2) and RFC 8009 §4 string-to-key
The RFC 8009 (SHA-2) and RC4-HMAC families live in the rfc8009 and rc4 modules.
Conformance
Known-answer tested, not just round-tripped:
nfold("012345", 64) == 0xbe072631276b1955(RFC 3961 §5.2)- PBKDF2-HMAC-SHA1 vs RFC 6070; HMAC-SHA1/MD5 vs RFC 2202; NT-hash + RC4 vs canonical vectors
- AES-SHA1 string-to-key vs RFC 3962; DR/DK determinism + AES
random_to_key = identity - AES-128/256 CTS round-trips across the CTS edge lengths (16 / 17 / 31 / 32 / 47 / 48 / 64 B)
- RFC 8009 §7 vectors — Kc/Ke/Ki derivation + all four sample encryptions, both etypes
- RC4-HMAC — byte-exact differential vs an independent reference, across key-usages × lengths
- encrypt-then-integrity round-trip + tamper / wrong-usage / truncation rejection
- the DER/message decoders fuzz-clean (coverage-guided, billions of iterations, zero panics)
Bringing up the RFC 8009 §7 known-answer vectors surfaced and fixed a latent CBC-CS3 edge case (the last-two-block swap for exact-multiple inputs) that round-trip tests alone did not catch.
cargo test → 65 passing (crypto + DER/message codec + client), plus one env-gated
live-KDC integration test that no-ops without credentials (so cargo test and CI stay
offline; no credentials or host identifiers live in the source).
Dependencies
aes, hmac, sha1, sha2, md4, md-5 (RustCrypto) and getrandom; RC4 is
hand-rolled. No FFI, no system krb5, #![forbid(unsafe_code)]. The differential
oracles (picky-krb, picky-asn1-der, ms-pac-forge) are dev-dependencies only.
Scope
kerbcore is the protocol core — encryption-type crypto, the wire codec, and a plain
AS/TGS client. It performs no network I/O (the caller owns the socket) and ships no
ticket-forging or attack tooling; those belong in higher-level tools built on top.
License
MIT.