quicproquo/docs/SECURITY-AUDIT.md

Security Audit

This document is a security audit of the quicprochat codebase as of the audit date. It aligns with the Threat Model and Production Readiness Audit. The project has not undergone a formal third-party audit; this is an internal review.

---

Executive Summary

Area	Finding	Severity
Authentication & sessions	Token comparison constant-time; session tokens CSPRNG; OPAQUE used correctly; no secrets in logs	✅ Strong
Cryptography	MLS, Ed25519, hybrid KEM, zeroization where appropriate; Argon2/ChaCha20 for state	✅ Strong
Transport (TLS)	TLS 1.3 only; client verifies server cert; self-signed default is documented weakness	⚠️ Known gap
Authorization (DS/AS)	Enqueue/fetch/fetchWait/key ops require auth + identity binding (or sealed_sender); health unauthenticated by design	✅ Appropriate
Input validation & limits	Key/recipient length, payload/KeyPackage size, queue depth, rate limit, UTF-8 username	✅ Good
Secrets handling	No tokens/keys/passwords in logs; DB key optional (documented); state encryption optional	✅ Good
Dependency hygiene	No cargo audit in tree; recommend adding and running in CI	⚠️ Recommendation

Overall: The design and implementation are security-conscious and match the documented threat model. Remaining risks are largely documented (self-signed TLS, metadata visibility, BasicCredential) or operational (deps, production config).

---

1. Authentication and Session Management

1.1 Token comparison

• Location: crates/quicprochat-server/src/auth.rs
• Finding: Bearer token and identity key comparisons use subtle::ConstantTimeEq (ct_eq). Length is checked before comparison where applicable.
• Status: ✅ No timing leakage from token or identity comparison.

1.2 Session token generation

• Location: crates/quicprochat-server/src/node_service/auth_ops.rs (login finish)
• Finding: Session tokens are 32 bytes from rand::RngCore::fill_bytes(&mut rand::rngs::OsRng, &mut token). Stored in sessions with TTL (24h). Expired sessions are removed on next use.
• Status: ✅ Cryptographically strong, single-use style (opaque 32-byte token).

1.3 OPAQUE (RFC 9497)

• Location: crates/quicprochat-core/src/opaque_auth.rs, server auth_ops.rs
• Finding: Shared OpaqueSuite (Ristretto255, Triple-DH, Argon2id). Server never sees password. Registration and login flows use ServerRegistration/ServerLogin correctly. Pending login state is stored server-side and removed on consume. Identity key is bound at login finish; mismatch returns E016 and is not logged with secrets.
• DoS: Pending-login check runs before expensive OPAQUE work (login start); repeated attempts for the same username within 60s are rejected early.
• Status: ✅ Correct usage; DoS mitigation in place.

1.4 Audit logging (no secrets)

• Finding: Comments in auth_ops.rs and delivery.rs explicitly forbid logging tokens, passwords, or raw keys. Logged fields: username, recipient/key prefix (fmt_hex), payload length, seq, counts. Login success/failure and rate-limit hit are logged without session token or identity.
• Status: ✅ No sensitive material in logs.

---

2. Cryptography

2.1 MLS and identity

• Location: quicprochat-core (group, identity, keypackage)
• Finding: MLS ciphersuite MLS_128_DHKEMX25519_AES128GCM_SHA256_Ed25519 (RFC 9420). Ed25519 identity seed stored in Zeroizing<[u8; 32]>; zeroize-on-drop. KeyPackages validated for ciphersuite before server stores. Single-use KeyPackage semantics enforced (consume-on-fetch).
• Status: ✅ Aligns with key lifecycle and zeroization goals.

2.2 Hybrid KEM (X25519 + ML-KEM-768)

• Location: crates/quicprochat-core/src/hybrid_kem.rs
• Finding: Hybrid keypair and shared secrets use Zeroizing where appropriate. HKDF domain separation (quicprochat-hybrid-v1). ChaCha20-Poly1305 for AEAD. Versioned envelope.
• Status: ✅ PQ-ready envelope layer; secret handling is careful.

2.3 Client state encryption (QPCE)

• Location: crates/quicprochat-client/src/client/state.rs
• Finding: Optional password protection: Argon2id (default params) for key derivation, ChaCha20-Poly1305, random salt and nonce. Derived key held in Zeroizing during use. Unencrypted state is a documented option (e.g. dev).
• Recommendation: Document Argon2 params (memory, iterations) for auditability; consider explicit Argon2::new() with named params in a future revision.
• Status: ✅ Appropriate for optional at-rest protection.

---

3. Transport (TLS / QUIC)

3.1 Server TLS

• Location: crates/quicprochat-server/src/tls.rs
• Finding: TLS 1.3 only. No client cert. ALPN capnp. When not in production, missing cert/key triggers self-signed generation; key file permissions set to 0o600 on Unix. Production mode requires existing cert/key (no auto-generation).
• Status: ✅ Matches documented design; self-signed limitation is documented in threat model.

3.2 Client TLS

• Location: crates/quicprochat-client/src/client/rpc.rs
• Finding: Client loads CA cert from file, builds RootCertStore with that single cert, uses it for server verification. Server name from CLI/env is used for connection (SNI and cert verification). No custom bypass.
• Status: ✅ Proper verification against provided CA; trust-on-first-use / self-signed caveat is documented.

---

4. Authorization and Access Control

4.1 RPC auth matrix

RPC	Auth required	Identity binding
health	No	N/A (liveness)
opaqueLoginStart/Finish, opaqueRegisterStart/Finish	No (password/session flow)	After login
uploadKeyPackage, fetchKeyPackage	Yes	Must match identity_key (or allow_insecure)
enqueue	Yes	Must match recipient_key unless sealed_sender
fetch, fetchWait	Yes	Must match recipient_key (or allow_insecure)
uploadHybridKey, fetchHybridKey	Yes	Must match identity_key (or allow_insecure)
publishEndpoint, resolveEndpoint	Yes	Publish: match identity_key; Resolve: any valid token

• Finding: Sensitive operations require validate_auth_context and, where relevant, require_identity_or_request. Fetch/fetchWait ensure the authenticated identity matches the requested recipient_key, so only the recipient (or someone with their session) can pull messages. With sealed_sender, enqueue only requires a valid token (no identity binding to sender).
• Status: ✅ Authorization consistent with design; sealed_sender trade-off is documented.

4.2 Rate limiting

• Location: crates/quicprochat-server/src/auth.rs, delivery.rs
• Finding: Per-token rate limit (e.g. 100 enqueues per 60s). Enqueue path checks before storage. Queue depth and payload size caps (1000 messages, 5 MB) enforced.
• Status: ✅ Limits in place to curb abuse and DoS.

---

5. Input Validation and Limits

• Identity/recipient keys: Rejected unless length exactly 32 bytes (E004).
• Payload: Non-empty; max 5 MB (E005, E006).
• KeyPackage: Non-empty; max 1 MB (E007, E008); ciphersuite validated before store (E021).
• Username: Non-empty; must be valid UTF-8 (E011, E020).
• Wire version: Rejected if > CURRENT_WIRE_VERSION (E012).
• Cap'n Proto: Server and client set traversal_limit_in_words(Some(4 * 1024 * 1024)) (4M words = 32 MiB) to bound parsing DoS.
• Status: ✅ Validation and limits are consistently applied.

---

6. Storage and Persistence

6.1 Server

• File-backed store: Mutex-protected; lock errors mapped to StorageError (no unwrap in hot path). OPAQUE server setup file permissions 0o600 on Unix.
• SQL store: Optional SQLCipher with db_key; empty key = plaintext (documented). Production validation requires non-empty db_key when backend is SQL. User records use INSERT (no OR REPLACE) and unique constraint; duplicate user returns StorageError::DuplicateUser (E018).
• Status: ✅ Matches production-readiness and auth design; DB encryption caveat documented.

6.2 Client

• State file: Optional QPCE encryption (Argon2id + ChaCha20-Poly1305). Unencrypted state contains identity seed; documented.
• Keystore: Persisted for HPKE init keys so Welcome can be processed after restart; path and format documented.
• Status: ✅ Acceptable for threat model; optional encryption and handling of secrets are clear.

---

7. Known Gaps (from Threat Model and Docs)

These remain as documented, not new findings:

1. Self-signed TLS: MITM possible on first connection if client does not pin or verify out-of-band. Mitigation: certificate pinning or CA-signed certs.
2. No client auth on DS (by design): Anyone with a valid token can enqueue to any recipient_key when identity is not required (e.g. sealed_sender). Rate limit and queue/payload caps mitigate abuse.
3. BasicCredential only: No CA or revocation; key substitution possible if AS is compromised. Mitigation: Key Transparency or X.509 credentials.
4. Metadata: Server sees recipient_key, timing, sizes; Sealed Sender and PIR are future mitigations.

---

8. Recommendations

8.1 High value

• Dependency audit: Run cargo install cargo-audit then cargo audit locally (and in CI if available) to check for known-vulnerable dependencies. Fix or document any findings. See Checking dependencies below.
• Argon2 params: Implemented: client state KDF now uses explicit Argon2id parameters (19 MiB memory, 2 iterations, 1 lane) in quicprochat-client so they are auditable.

8.2 Medium value

• Certificate pinning: To pin the server, use the server's certificate as the client's ca_cert (e.g. copy server-cert.der from the server and pass it via --ca-cert or QPC_CA_CERT). Do not use a general CA unless you intend to trust that CA for all servers. See Certificate pinning below.
• Health endpoint: The health RPC is unauthenticated by design for liveness probes and load balancers; this is documented in code and in this audit.

8.3 Lower priority

• Cap'n Proto traversal limit: Implemented: reduced to 4M words (32 MiB) with a named constant; trade-off documented in code.
• Username enumeration: OPAQUE login start uses get_user_record; timing or response shape might still reveal user existence. Mitigation in scope: consider constant-time or uniform response for unknown users in a future revision (e.g. fixed dummy work when user not found) without weakening OPAQUE. Current code does not implement this.

---

9. Summary Table

Category	Item	Status
Auth	Constant-time token/identity comparison	✅
Auth	Session token from CSPRNG, not logged	✅
Auth	OPAQUE used correctly; no password on server	✅
Auth	Pending-login DoS check before OPAQUE work	✅
Auth	No secrets in audit logs	✅
Crypto	MLS ciphersuite, KeyPackage validation	✅
Crypto	Identity seed zeroized on drop	✅
Crypto	Hybrid KEM and state encryption use Zeroizing	✅
Transport	TLS 1.3 only; client verifies server cert	✅
Transport	Self-signed default (documented weakness)	⚠️ Known
Authz	Enqueue/fetch/key/p2p require auth + identity where required	✅
Authz	Rate limit and size/depth limits	✅
Input	Lengths, sizes, UTF-8, wire version, traversal limit	✅
Storage	Server: lock Result, DB key in prod, duplicate user	✅
Storage	Client: optional QPCE; explicit Argon2 params; unencrypted state documented	✅
Deps	Run cargo audit locally/CI (see Checking dependencies)	⚠️ Recommend

---

Checking dependencies

To check for known vulnerabilities in dependencies:

cargo install cargo-audit
cargo audit

Fix or document any reported issues. Running cargo audit in CI (e.g. GitHub Actions) is recommended.

---

Certificate pinning

The client trusts the server certificate(s) in the file given by --ca-cert (or QPC_CA_CERT). To pin a specific server:

1. Obtain the server's certificate (e.g. copy data/server-cert.der from the server, or export from your deployment).
2. Use that file as the client's ca_cert. The client will only connect to a server that presents that exact certificate (or chain).
3. Do not use a broad CA bundle as ca_cert unless you intend to trust any server certified by that CA.

This gives trust-on-first-use behaviour when you deploy the server and then distribute its cert to clients.

---

Related Documents

• Threat Model
• Production Readiness Audit
• Cryptography Overview
• Key Lifecycle and Zeroization