Use GitHub private vulnerability reporting. Do not post exploit details, credentials, private SDP addresses or media in a public issue. Include the affected version/commit, a minimal local reproduction and impact. There is no guaranteed response time; maintainers will investigate and coordinate a fix.

The latest 0.x line is experimental and receives fixes as development proceeds. Older development packages are not maintained separately. No production-readiness or independent security-audit claim is made.

The project has not received an independent security audit. Parser and cryptographic verification tests do not establish complete WebRTC transport security.

Current parsers validate datagram boundaries before exposing zero-copy spans. They do not open sockets or resolve names. A successful RTP parse says nothing about packet authentication. Future transport code must authenticate packets before delivering media. The caller must keep backing buffers unchanged while a parsed view or attribute enumerator is in use; read-only spans do not make the underlying storage immutable.

The assembly is public-signed with the repository's public strong-name key for stable assembly identity. This is not a publisher-authenticity signature or a security boundary. No private signing key is committed.

STUN MESSAGE-INTEGRITY uses the HMAC-SHA1 wire algorithm required for existing ICE interoperability. Hashing is performed by .NET cryptography, and integrity comparisons use CryptographicOperations.FixedTimeEquals. The caller must supply the correctly derived credential key. The owned TURN allocation implements bounded ASCII long-term credential derivation and negotiated SHA256/legacy integrity; see TURN scope. FINGERPRINT is a CRC for protocol identification and corruption detection, not authentication.

SRTP/SRTCP contexts authenticate before returning decrypted packets, enforce replay windows and refuse sender index reuse. They require fresh directional key generations; standalone contexts do not authenticate peers or negotiate keys. See SRTP/SRTCP scope. The independent peer and synthetic vectors establish profile interoperability, not an independent security audit. DtlsSrtpTransport authenticates both peer fingerprints, verifies handshake signatures/Finished and derives directional SRTP material. Signaled fingerprints must come from trusted signaling. DTLS supports only a bounded DTLS 1.2/EMS/P-256/AES-GCM subset; unsupported negotiations fail closed. See DTLS scope and key limits.

No untrusted packet may trigger unbounded allocation, reassembly, retry loops or network destinations. Add explicit limits when each stateful transport component is introduced. SDP-provided ICE servers, DNS answers, candidates and datagrams remain untrusted input. The initial peer owns all transports and closes them on establishment cancellation or transport failure. Only authenticated/replay-checked SRTP reaches Opus routing; wrong MID/payload/direction, conflicting/local sources and excess source admission are rejected. Known RTP extension framing is checked before routing. RTCP framing checks do not constitute semantic feedback/SDES validation. Bounded drop-oldest media queues report losses and add no intentional jitter delay. See peer scope. The SDP parser bounds UTF-8 bytes, lines, media sections, attributes, candidates, codecs and extensions. It performs no network/DNS work. Initial Opus/data negotiation validates accepted media, BUNDLE credentials, fingerprint, DTLS roles and extension directions before the caller starts transport. Trusted signaling must authenticate those credentials and fingerprints. Parsing a hostname does not authorize resolving it; candidate destination policy remains the transport caller's responsibility. SDES/plaintext keys are rejected. Unsupported active media fails negotiation. See SDP limits; these checks are not a complete JSEP/security audit.

SCTP validates CRC32C, verification tags, framing and bounded state over DTLS. DCEP channels have aggregate byte/count admission and strict UTF-8/control validation. PR-SCTP must be negotiated; FORWARD-TSN advances and stream entries are bounded and validated before receive-state changes. Abandonment releases payload storage but retains bounded metadata until acknowledgment. See SCTP/data-channel scope for limits and missing features. CRC32C is not authentication. Observe channel-owner failures and close the eventual peer connection. Stream reset is negotiated and validates parameter framing, unique stream IDs, request sequence and bounded TSN barriers before changing state. IDs remain reserved until both directions finish; invalid DCEP still terminates its owner.

UDP relay integration binds authenticated ICE transactions, early checks, nomination, consent and selected ciphertext to the exact local path and remote endpoint. TURN IP permissions are routing admission, not peer authentication. A failed selected relay closes its owner rather than moving authentication to another path. Relay-only excludes host connectivity; SDP related mapped addresses can still disclose address metadata. See ICE path scope.

General negotiation, complete NAT traversal and real consumer acceptance remain unimplemented. Consumers must not migrate on DTLS/SRTP interoperability alone.

TURN TCP/TLS stream framing and write admission have fixed bounds. An interrupted write closes its connection because the remote side might have a partial frame. No authentication state is reused through reconnect or fallback. TLS uses platform SslStream, TLS 1.2/1.3, strict target-name/chain/time/server-purpose validation and no validation callback. Custom DER roots replace system trust only when explicitly supplied; they are snapshotted before connection. Revocation defaults to Online and may contact certificate revocation services; private isolated fixtures explicitly choose NoCheck. AIA issuer downloads are disabled, so servers must supply their intermediate chain. This is transport interoperability, not an independent security audit.