]> Vulture Labs

calin@vulturelabs.com

Internet Independent Submission This document specifies Pilot Protocol, an overlay network that provides autonomous AI agents with virtual addresses, port-based service multiplexing, reliable and unreliable transport, NAT traversal, encrypted tunnels, and a bilateral trust model. Pilot Protocol operates as a network and transport layer beneath application-layer agent protocols such as A2A and MCP. It encapsulates virtual packets in UDP datagrams for transit over the existing Internet. The protocol gives agents first-class network citizenship --- stable identities, reachable addresses, and standard transport primitives --- independent of their underlying network infrastructure.

Introduction AI agents are autonomous software entities that reason, plan, and execute tasks. As agents become more prevalent, they need to communicate with each other across heterogeneous network environments: cloud, edge, behind NAT, and across organizational boundaries. Current agent protocols (MCP, A2A) operate at the application layer over HTTP, assuming agents have stable, reachable endpoints. This assumption fails for a large class of deployments. Pilot Protocol is an overlay network stack that gives agents network-layer primitives: virtual addresses, ports, reliable streams, unreliable datagrams, NAT traversal, encrypted tunnels, name resolution, and a bilateral trust model. It is positioned as the network/transport layer beneath application-layer agent protocols --- analogous to how TCP/IP sits beneath HTTP.

Design Principles The protocol is designed around five principles: Agents are first-class network citizens. Every agent gets a unique virtual address, can bind ports, listen for connections, and be reached by any authorized peer. The network boundary is the trust boundary. Network membership serves as the primary access control mechanism. Joining a network requires meeting its rules. Transport agnosticism. The protocol provides reliable streams (TCP-equivalent) and unreliable datagrams (UDP-equivalent). Anything that runs on TCP/IP can run on the overlay. Minimize the protocol, maximize the surface. The protocol defines addressing, packets, and transport. Application-level message formats are layers built on top. Practical over pure. The protocol uses a centralized registry for address assignment and a centralized beacon for NAT traversal. Full decentralization is a future goal, not a prerequisite.

Relationship to Existing Protocols Pilot Protocol operates at the network and transport layers of the overlay stack. It is complementary to, not competitive with, application-layer agent protocols: A2A defines what agents say to each other. Pilot defines how they reach each other. MCP defines agent-to-tool interfaces. Pilot provides the transport substrate. QUIC is a potential underlay transport. Pilot could run over QUIC instead of raw UDP. LISP provides conceptual precedent for identity/locator separation. VXLAN and GENEVE are overlay encapsulation precedents at the data link layer. Pilot operates at the network layer.

Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Agent:

An autonomous software entity capable of reasoning, planning, and executing tasks without continuous human supervision.

Daemon:

The local Pilot Protocol process that implements the virtual network stack. It maintains a UDP tunnel, handles routing, session management, and encryption. Analogous to a virtual NIC.

Driver:

An SDK or library that agents import to communicate with the local daemon over IPC. Provides the application-facing API (listen, dial, read, write, close).

Registry:

A centralized service that assigns virtual addresses, maintains the address-to-locator mapping table, manages network membership, and stores public keys.

Beacon:

A service that provides NAT traversal coordination: endpoint discovery (STUN-like), hole-punch coordination, and relay fallback.

Virtual Address:

A 48-bit overlay address assigned to an agent, independent of its underlying IP address.

Trust Pair:

A bilateral trust relationship between two agents, established through explicit mutual consent.

Architecture Overview

Protocol Stack Pilot Protocol is a five-layer overlay stack: Layer Function Application HTTP, RPC, custom protocols (above Pilot) Session Reliable streams, unreliable datagrams Network Virtual addresses, ports, routing Tunnel NAT traversal, UDP encapsulation, encryption Physical Real Internet (IP/TCP/UDP) The overlay handles addressing, routing, and session management. The underlying Internet handles physical delivery.

Component Roles

Registry:

Assigns virtual addresses, maintains address table, manages networks and trust pairs, relays handshake requests for private nodes. Runs on TCP. The only globally reachable component.

Beacon:

Provides STUN-like endpoint discovery, hole-punch coordination, and relay fallback for symmetric NAT. Runs on UDP.

Daemon:

Core protocol implementation running on each participating machine. Maintains a single UDP socket, multiplexes all virtual connections, handles tunnel encryption, and exposes a local IPC socket for drivers.

Driver:

Client SDK that agents import. Connects to the local daemon via Unix domain socket. Implements standard network interfaces (listeners, connections).

Nameserver:

DNS equivalent for the overlay. Runs as a service on virtual port 53, resolving human-readable names to virtual addresses.

Gateway:

Bridge between the overlay and standard IP. Maps virtual addresses to local IPs, allowing unmodified TCP programs to reach agents.

Addressing

Virtual Address Format Addresses are 48 bits, split into two fields:

Network ID (16 bits):

Identifies the network or topic. Network 0 is the global backbone; all nodes are members by default. Networks 1-65534 are created for specific purposes. Network 65535 is reserved.

Node ID (32 bits):

Identifies the individual agent within a network. Supports over 4 billion nodes per network.

Text Representation Addresses are written as N:XXXX.YYYY.ZZZZ where: N is the network ID in decimal XXXX.YYYY.ZZZZ is the node ID as three groups of 4 hexadecimal digits Examples: 0:0000.0000.0001 --- Node 1 on the backbone 1:00A3.F291.0004 --- A node on network 1

Socket Addresses A socket address appends a port: 1:00A3.F291.0004:1000

Special Addresses Address Meaning 0:0000.0000.0000 Unspecified / wildcard 0:0000.0000.0001 Registry 0:0000.0000.0002 Beacon 0:0000.0000.0003 Nameserver X:FFFF.FFFF.FFFF Broadcast on network X

Ports

Port Ranges Virtual ports are 16-bit unsigned integers (0-65535): Range Purpose 0-1023 Reserved / well-known 1024-49151 Registered services 49152-65535 Ephemeral / dynamic

Well-Known Ports Port Service Description 0 Ping Liveness checks 1 Control Daemon-to-daemon control 7 Echo Echo service (testing) 53 Name resolution Nameserver queries 80 Agent HTTP Web endpoints 443 Secure channel X25519 + AES-256-GCM 444 Trust handshake Peer trust negotiation 1000 Standard I/O Text stream between agents 1001 Data exchange Typed frames (text, binary, JSON, file) 1002 Event stream Pub/sub with topic filtering 1003 Task submission Task lifecycle and reputation scoring

Packet Format

Header Layout The fixed packet header is 34 bytes:

All multi-byte fields are in network byte order (big-endian).

Field Definitions Field Offset Size Description Version 0 4 bits Protocol version. Current: 1 Flags 0 4 bits SYN (0x1), ACK (0x2), FIN (0x4), RST (0x8) Protocol 1 1 byte Transport type (see ) Payload Length 2 2 bytes Payload length in bytes (max 65535) Source Network 4 2 bytes Source network ID Source Node 6 4 bytes Source node ID Dest Network 10 2 bytes Destination network ID Dest Node 12 4 bytes Destination node ID Source Port 16 2 bytes Source port Dest Port 18 2 bytes Destination port Sequence Number 20 4 bytes Byte offset of this segment Ack Number 24 4 bytes Next expected byte from peer Window 28 2 bytes Advertised receive window in segments Checksum 30 4 bytes CRC32 over header + payload

Protocol Types Value Name Description 0x01 Stream Reliable, ordered delivery (TCP-like) 0x02 Datagram Unreliable, unordered (UDP-like) 0x03 Control Internal control messages

Flag Definitions Bit Name Description 0 SYN Synchronize --- initiate connection 1 ACK Acknowledge --- confirm receipt 2 FIN Finish --- close connection 3 RST Reset --- abort connection

Checksum Calculation The checksum is computed as follows: Set the 4-byte checksum field to zero. Compute CRC32 (IEEE polynomial) over the entire header (34 bytes with zeroed checksum field) concatenated with the payload bytes. Write the resulting 32-bit value into the checksum field in big-endian byte order. Receivers MUST verify the checksum and discard packets with incorrect values. Note: CRC32 detects accidental corruption but does not provide cryptographic integrity. Tamper resistance is provided by tunnel-layer encryption ().

Tunnel Encapsulation Virtual packets are encapsulated in UDP datagrams for transit over the real Internet. Four frame types are defined, distinguished by a 4-byte magic value.

Plaintext Frame (PILT)

The magic bytes 0x50494C54 ("PILT") indicate an unencrypted Pilot Protocol frame. The header and payload follow immediately.

Encrypted Frame (PILS)

The magic bytes 0x50494C53 ("PILS") indicate an encrypted frame.

SenderID:

4-byte Node ID of the sending daemon, in big-endian. Used by the receiver to look up the shared AES-256-GCM key for this peer.

Nonce:

12-byte GCM nonce. See for construction.

Ciphertext:

The Pilot Protocol header and payload, encrypted with AES-256-GCM . The ciphertext includes a 16-byte authentication tag appended by GCM.

The encryption key is derived from an X25519 ECDH exchange between the two daemons (see ).

Key Exchange Frame (PILK)

Anonymous key exchange. The sender transmits its ephemeral X25519 public key (32 bytes, per ). Both sides compute the ECDH shared secret and derive an AES-256-GCM key. PILK provides confidentiality but not authentication. An active attacker can perform a man-in-the-middle attack by substituting their own public key. See for the authenticated variant.

Authenticated Key Exchange Frame (PILA)

Authenticated key exchange. In addition to the X25519 public key, the sender includes its Ed25519 public key (32 bytes, per ) and a 64-byte Ed25519 signature. The signature covers the concatenation of: The ASCII string "auth" (4 bytes) The sender's Node ID (4 bytes, big-endian) The X25519 public key (32 bytes) The receiver verifies the signature using the sender's Ed25519 public key, which it obtains from the registry and cross-checks against the claimed Node ID. This binds the ephemeral X25519 key to the sender's persistent identity, preventing man-in-the-middle attacks. Daemons with persistent Ed25519 identities SHOULD use PILA. Daemons without persistent identities fall back to PILK.

Session Layer

Connection State Machine

SYN_SENT ^ | | SYN-ACK recv | | | v | ESTABLISHED <--- Listen() + SYN recv | | | Close() | | | v | FIN_WAIT | | | ACK recv | | | v | TIME_WAIT | | | 10s timeout | | +--------------------+ ]]>

Three-Way Handshake Connection establishment uses a three-way handshake:

| | | |<--- SYN+ACK seq=Y ack=X+1 -----| | | |-------- ACK ack=Y+1 ----------->| | | | ESTABLISHED | ESTABLISHED ]]>

The initiator selects an initial sequence number X. The responder selects its own initial sequence number Y and acknowledges X+1. The initiator confirms by acknowledging Y+1. Both sides include their advertised receive window in the Window field of the SYN and SYN-ACK packets.

Connection Teardown

| | | |<------- ACK ack=N+1 -----------| | | | TIME_WAIT (10s) | CLOSED | | | CLOSED | ]]>

The closer sends FIN, waits for ACK, and enters TIME_WAIT for 10 seconds. The 10-second TIME_WAIT is shorter than TCP's typical 2*MSL because overlay RTTs are bounded by the underlay network.

Sequence Number Arithmetic Sequence numbers are 32-bit unsigned integers with wrapping comparison per :

0 ]]>

This correctly handles wraparound at 2^32.

Reliable Delivery The Stream protocol (0x01) provides reliable, ordered byte stream delivery using a sliding window mechanism.

Retransmission Timeout (RTO) RTO is computed per : SRTT (Smoothed RTT): updated with alpha = 1/8 RTTVAR (RTT Variance): updated with beta = 1/4 RTO = SRTT + max(G, 4 * RTTVAR) G (clock granularity floor) = 10 ms RTO is clamped to the range 200 ms to 10 s SYN packets are retransmitted with exponential backoff: 1s, 2s, 4s, 8s, up to 5 retries. Data segments allow up to 8 retransmission attempts before the connection is closed.

Out-of-Order Handling Segments received out of order are buffered and delivered to the application in sequence order when gaps are filled.

Selective Acknowledgment (SACK) When the receiver has out-of-order segments, it encodes SACK blocks in the ACK payload. Each SACK block is a pair of 32-bit sequence numbers representing a contiguous range of received bytes beyond the cumulative ACK point. Up to 4 SACK blocks are encoded per ACK. The sender uses SACK information to retransmit only the missing segments, skipping segments the peer has already received.

Congestion Control The protocol implements TCP-style congestion control:

Slow Start:

The congestion window (cwnd) starts at 10 segments (40 KB) per and grows by one segment for each ACK received, until cwnd reaches the slow-start threshold (ssthresh).

Congestion Avoidance:

After cwnd reaches ssthresh, growth switches to additive-increase: cwnd grows by approximately one segment per round-trip time (Appropriate Byte Counting per ).

Fast Retransmit:

After 3 duplicate pure ACKs (data packets with piggybacked ACKs are excluded per Section 3.2), the sender retransmits the missing segment without waiting for RTO.

Multiplicative Decrease:

On loss detection (fast retransmit or RTO), ssthresh is set to max(cwnd/2, 2 segments). On RTO, cwnd is additionally reset to 1 segment (Tahoe behavior).

The maximum congestion window is 1 MB. The maximum segment size (MSS) is 4096 bytes.

Flow Control Each ACK carries the receiver's advertised window --- the number of free segments in its receive buffer. The sender's effective window is:

This prevents a fast sender from overwhelming a slow receiver.

Zero-Window Probing When the receiver advertises a zero window, the sender enters persist mode and sends 1-byte probe segments at exponentially increasing intervals until the receiver opens its window.

Write Coalescing (Nagle's Algorithm) Small writes are buffered when unacknowledged data is in flight and flushed when: The buffer reaches MSS (4096 bytes), or All previous data is acknowledged, or A 40 ms timeout expires. This reduces packet overhead for applications performing many small writes. The algorithm can be disabled per-connection with a NoDelay option, analogous to TCP_NODELAY.

Automatic Segmentation Large writes are automatically segmented into MSS-sized chunks (4096 bytes) by the daemon. Applications can write arbitrarily large buffers without manual chunking.

Delayed ACKs Instead of sending an ACK for every received segment, the daemon batches up to 2 segments or 40 ms (whichever comes first). When out-of-order data is present, ACKs are sent immediately with SACK blocks to trigger fast retransmit. When data is sent on a connection, the pending delayed ACK is cancelled because the outgoing data packet piggybacks the latest cumulative ACK and receive window.

Keepalive and Idle Timeout Keepalive probes (empty ACKs) are sent every 30 seconds to idle connections. Connections idle for 120 seconds are automatically closed. These timers are appropriate for the overlay's use case (agent communication), where stale connections should be reclaimed promptly.

TIME_WAIT Closed connections enter TIME_WAIT for 10 seconds before being removed. During TIME_WAIT, the connection occupies its port binding (preventing reuse confusion with delayed packets) but does not count as active.

NAT Traversal

STUN-Based Endpoint Discovery On startup, the daemon sends a UDP probe to the beacon. The beacon observes the daemon's public IP address and port (as mapped by NAT) and reports it back. This follows the mechanism described in (Session Traversal Utilities for NAT). The discovered public endpoint is registered with the registry as the daemon's locator. For daemons with known public endpoints (e.g., cloud VMs), the -endpoint host:port flag skips STUN and registers the specified endpoint directly.

Hole Punching When daemon A wants to reach daemon B and both are behind NAT: Daemon A sends a punch request to the beacon, specifying B's Node ID. The beacon looks up B's registered endpoint and sends a punch command to both A and B, instructing each to send a UDP packet to the other's observed endpoint. Both daemons send UDP packets to each other simultaneously, punching holes in their respective NATs. Subsequent packets flow directly between A and B. This works for Full Cone, Restricted Cone, and Port-Restricted Cone NAT types.

Relay Fallback When hole punching fails (typically Symmetric NAT, where the mapped port changes per destination), the beacon provides transparent relay:

| Beacon | ------> | Daemon B | +----------+ relay +----------+ relay +----------+ ]]>

The relay frame format:

The beacon unwraps the relay header and forwards the payload to the destination daemon. Relay is transparent to the session layer --- the virtual packet inside the relay frame is identical to a directly-delivered packet.

Connection Establishment Strategy When dialing a remote daemon, the connection strategy is: Attempt 3 direct UDP sends to the peer's registered endpoint. If all 3 fail, switch to relay mode through the beacon. Attempt 3 relay sends. If all relay attempts fail, return an error to the application. The switch from direct to relay is automatic and transparent to the application layer.

Security

Identity Each node receives an Ed25519 keypair from the registry upon registration. The private key serves as the node's identity credential. The registry holds all public keys and can verify signatures. Identities may be persisted to disk so that a node retains its keypair and virtual address across restarts. On restart with a persisted identity, the daemon re-registers with the stored public key and the registry restores the node's address and memberships.

Tunnel-Layer Encryption Tunnel encryption is enabled by default. On startup, each daemon generates an ephemeral X25519 keypair. When two daemons first communicate, they exchange public keys via PILK () or PILA () frames, compute an ECDH shared secret, and establish an AES-256-GCM cipher. All subsequent packets between the pair are encrypted (PILS frames), regardless of virtual port. The encryption is at the tunnel layer --- it protects all overlay traffic between two daemons, including connection handshakes. Tunnel encryption is backward-compatible: if a peer does not respond to key exchange, communication falls back to plaintext (PILT frames).

Authenticated Key Exchange Upgrade When a daemon has a persisted Ed25519 identity, the key exchange is upgraded from PILK to PILA (see ). The Ed25519 signature binds the ephemeral X25519 key to the node's persistent identity, preventing man-in-the-middle attacks. Implementations SHOULD use PILA when an Ed25519 identity is available.

Application-Layer Encryption (Port 443) Virtual port 443 provides end-to-end encryption between two agents, on top of any tunnel-layer encryption. The agents perform an X25519 ECDH handshake to derive a shared secret, then use AES-256-GCM for all subsequent data. Each encrypted frame:

This provides defense in depth: even if the tunnel encryption is compromised (e.g., by a compromised intermediate daemon in a future multi-hop topology), port 443 data remains protected.

Trust Handshake Protocol (Port 444) Port 444 implements a bilateral trust negotiation protocol. Two agents exchange trust requests with justification strings and must both approve before a trust relationship is established. Three auto-approval paths exist: Mutual handshake: If both agents independently request trust with each other, the relationship is auto-approved. Network trust: If both agents share a non-backbone network, the relationship is auto-approved (network membership serves as a trust signal). Manual approval: If neither condition is met, the request is queued for the receiving agent's operator to approve or reject. Trust pairs are recorded in the registry and persist across restarts. Trust is revocable: revoking trust immediately prevents further communication.

Privacy Model Agents are private by default: A node's physical IP:port is never disclosed in registry responses unless the node has explicitly opted into public visibility. Resolving a private node's endpoint requires one of: (a) the node is public, (b) a mutual trust pair exists, or (c) both nodes share a non-backbone network. Listing nodes on the backbone (network 0) is rejected by the registry. Non-backbone networks allow listing since membership is the trust boundary.

Rate Limiting The registry enforces per-connection sliding window rate limits using a token-bucket algorithm with per-source tracking. Clients that exceed the limit receive throttle responses. Daemons implement SYN rate limiting to mitigate connection flood attacks.

IPC Security The daemon's Unix domain socket is created with mode 0600, restricting access to the socket owner. This prevents unprivileged processes on the same machine from issuing commands to the daemon.

Nonce Management

Construction AES-256-GCM nonces are 96 bits (12 bytes), constructed as:

Prefix:

4 bytes generated from a cryptographically secure random source when the tunnel session is established. Unique per session with overwhelming probability.

Counter:

8-byte unsigned integer, starting at 0, incremented by 1 for each packet encrypted. The counter MUST NOT be reset within a session.

Session Lifecycle A new tunnel session is established when two daemons perform an X25519 key exchange (PILK or PILA). Each session produces: A fresh AES-256-GCM key (from the ECDH shared secret) A fresh random nonce prefix Since each session uses a different key, nonces from different sessions cannot collide (different keys are independent encryption contexts).

Counter Exhaustion The 8-byte counter supports 2^64 encryptions per session. Implementations MUST re-key (initiate a new key exchange) before the counter reaches 2^64 - 1. In practice, at 1 million packets per second, counter exhaustion would take over 584,000 years.

Application-Layer Nonces (Port 443) Secure connections on port 443 use a separate nonce scheme: a monotonically increasing 8-byte counter zero-padded to 12 bytes. Each connection has an independent counter and key derived from its own X25519 handshake.

Version Negotiation

Version Field The 4-bit Version field in the packet header identifies the protocol version. The current version is 1. Version 0 is reserved and MUST NOT be used.

Handling Mismatches The initiator includes its protocol version in the SYN packet. The responder checks the version: If supported: echoes the same version in SYN-ACK. Both sides use this version for the connection's lifetime. If unsupported: sends RST. No version downgrade negotiation occurs. For non-SYN packets, if the Version field does not match the connection's established version, the packet is silently discarded. Implementations SHOULD log such events at debug level.

Future Extensibility Future protocol versions MAY extend the header format. Implementations MUST NOT assume a fixed header size based solely on the Version field --- they SHOULD use the version to determine the expected header layout.

Path MTU Considerations

Encapsulation Overhead The total per-packet overhead for encrypted tunnel frames is: Component Size PILS magic 4 bytes Sender Node ID 4 bytes GCM nonce 12 bytes Pilot header 34 bytes GCM authentication tag 16 bytes Total 70 bytes For plaintext frames (PILT), overhead is 38 bytes (4-byte magic + 34-byte header).

Effective Payload With a typical 1500-byte Ethernet MTU, 20-byte IP header, and 8-byte UDP header: Available for Pilot: 1500 - 28 = 1472 bytes Encrypted payload capacity: 1472 - 70 = 1402 bytes Plaintext payload capacity: 1472 - 38 = 1434 bytes

MSS Selection The default MSS of 4096 bytes exceeds single-packet capacity on standard Ethernet paths. Full-MSS segments will be fragmented into 3 IP fragments. This is acceptable on most networks but may fail on paths that block IP fragmentation. Recommendations: For Internet-facing deployments where IP fragmentation may be blocked, an MSS of 1400 bytes avoids fragmentation on virtually all paths. For datacenter or local deployments (jumbo frames), the default 4096 MSS is appropriate. Implementations SHOULD provide a configurable MSS option. Implementations SHOULD NOT set the Don't Fragment (DF) bit on UDP datagrams, allowing IP-layer fragmentation as a fallback.

Built-in Services

Echo (Port 7) The echo service reflects any data received back to the sender. It is used for liveness testing (ping) and throughput benchmarking.

Data Exchange (Port 1001) A typed frame protocol for structured data. Each frame carries a 4-byte type tag and a 4-byte length prefix: Type Value Description Text 0x01 UTF-8 text Binary 0x02 Raw bytes JSON 0x03 JSON document File 0x04 File with name metadata

Event Stream (Port 1002) A publish/subscribe broker. Agents subscribe to named topics and receive events published by any peer. Wildcard subscriptions (*) match all topics. The wire protocol uses newline-delimited text commands: SUB <topic> --- subscribe to a topic PUB <topic> <payload> --- publish an event EVENT <topic> <payload> --- delivered event (broker to subscriber)

Task Submission (Port 1003) A task lifecycle protocol. Agents submit tasks with descriptions, workers accept or decline, execute, and return results. A reputation score (polo score) adjusts based on execution efficiency.

IPC Protocol

Framing The daemon and driver communicate over a Unix domain socket using length-prefixed messages:

Maximum message size: 1,048,576 bytes (1 MB).

Command Set Cmd Name Direction Description 0x01 Bind Driver -> Daemon Bind a virtual port 0x02 BindOK Daemon -> Driver Confirm port binding 0x03 Dial Driver -> Daemon Connect to remote agent 0x04 DialOK Daemon -> Driver Connection established 0x05 Accept Daemon -> Driver Incoming connection 0x06 Send Driver -> Daemon Send data on connection 0x07 Recv Daemon -> Driver Receive data 0x08 Close Driver -> Daemon Close connection 0x09 CloseOK Daemon -> Driver Connection closed 0x0A Error Daemon -> Driver Error response 0x0B SendTo Driver -> Daemon Send datagram 0x0C RecvFrom Daemon -> Driver Receive datagram 0x0D Info Driver -> Daemon Query daemon status 0x0E InfoOK Daemon -> Driver Status response (JSON) 0x0F Handshake Driver -> Daemon Trust handshake command 0x10 HandshakeOK Daemon -> Driver Handshake result (JSON)

Security Considerations

CRC32 Limitations The packet checksum uses CRC32 (IEEE polynomial), which detects accidental corruption but provides no cryptographic integrity. An attacker who can modify packets in transit can recompute a valid CRC32. Integrity against active attackers is provided by tunnel-layer AES-256-GCM encryption, which MUST be used for all Internet-facing deployments.

Anonymous Key Exchange The PILK key exchange frame provides no identity binding. An active man-in-the-middle attacker can substitute their own X25519 public key, establishing separate encrypted sessions with each peer. The PILA authenticated key exchange () prevents this by binding the ephemeral key to an Ed25519 identity. Implementations SHOULD use PILA whenever an Ed25519 identity is available.

Registry as Trusted Third Party The registry is a centralized trusted third party. Compromise of the registry could allow: Address hijacking (reassigning a node's virtual address) Locator spoofing (returning incorrect IP:port for a node) Public key substitution (enabling identity impersonation) Metadata harvesting (enumerating registered nodes) Mitigations include TLS transport for registry connections, admin token authentication for write operations, and hot-standby replication for availability. Future work should explore distributed registry designs with consensus-based replication.

GCM Nonce Uniqueness AES-256-GCM security depends critically on nonce uniqueness under the same key. The nonce construction () guarantees uniqueness through a random prefix (unique per session) and a monotonic counter (never reset within a session). Since each key exchange produces a new key, nonces from different sessions are in independent cryptographic contexts. Implementations MUST NOT reuse nonces. Implementations MUST NOT reset the counter within a session. Implementations MUST re-key before counter exhaustion.

Metadata Exposure Even with tunnel encryption (PILS), the sender's Node ID is transmitted in cleartext (it is needed for the receiver to look up the decryption key). This allows a passive observer to determine which daemons are communicating, though the content and virtual addressing within the encrypted payload remain confidential.

Double Congestion Control Pilot Protocol implements congestion control at the overlay layer, while the underlay UDP-over-IP path may also be subject to network-level congestion signals (ICMP source quench, ECN). The overlay congestion control operates independently, which may lead to suboptimal behavior on heavily congested paths. This is a known issue shared with all overlay transport protocols.

Replay Protection Tunnel-layer AES-256-GCM provides implicit replay protection: GCM authentication will fail for replayed packets if the receiver tracks seen nonces. However, the current specification does not mandate a replay window. Implementations SHOULD track recently seen nonces and discard duplicates.

IPC as Trust Boundary The Unix domain socket IPC between daemon and driver is a trust boundary. The daemon trusts that any process connecting to the socket is authorized (enforced by filesystem permissions, mode 0600). If an attacker gains access to the socket, they can impersonate the local agent. Deployments SHOULD ensure the daemon runs under a dedicated user account.

IANA Considerations

Pilot Protocol Tunnel Magic Values This document requests the creation of a "Pilot Protocol Tunnel Magic Values" registry with the following initial entries: Magic Hex Description PILT 0x50494C54 Plaintext frame PILS 0x50494C53 Encrypted frame PILK 0x50494C4B Key exchange frame PILA 0x50494C41 Authenticated key exchange frame

Pilot Protocol Type Values This document requests the creation of a "Pilot Protocol Type Values" registry with the following initial entries: Value Name Description 0x01 Stream Reliable, ordered delivery 0x02 Datagram Unreliable, unordered delivery 0x03 Control Internal control messages

Pilot Protocol Well-Known Ports This document requests the creation of a "Pilot Protocol Well-Known Ports" registry with the following initial entries: Port Service Description 0 Ping Liveness checks 1 Control Daemon-to-daemon control 7 Echo Echo service 53 Name Resolution Nameserver 80 Agent HTTP Web endpoints 443 Secure End-to-end encrypted channel 444 Trust Trust handshake protocol 1000 StdIO Text stream 1001 DataExchange Typed frame protocol 1002 EventStream Pub/sub broker 1003 TaskSubmit Task lifecycle

Implementation Status Per , this section documents the known implementations of Pilot Protocol at the time of writing.

Go Reference Implementation

Organization:

Vulture Labs

Description:

Complete implementation of Pilot Protocol including daemon, driver SDK, registry, beacon, nameserver, gateway, and CLI (pilotctl). Implemented in Go with zero external dependencies.

Level of maturity:

Production-ready for experimental deployments.

Coverage:

All features specified in this document are implemented, including tunnel encryption (PILK/PILA/PILS), SACK, congestion control, flow control, Nagle's algorithm, automatic segmentation, NAT traversal (STUN, hole-punch, relay), trust handshake protocol, privacy model, and all built-in services.

Testing:

226+ tests (202 PASS, 24 SKIP). Integration tests validated across 5 GCP regions (US Central, US East, Europe West, US West, Asia East) with both public-IP and NAT-only topologies.

Licensing:

Proprietary.

Contact:

calin@vulturelabs.com

Python SDK

Organization:

Vulture Labs

Description:

Python client SDK using ctypes FFI to the Go shared library. Published on PyPI as pilotprotocol.

Level of maturity:

Beta.

Coverage:

Driver operations (dial, listen, accept, send, receive, close), datagram support, info queries.

Licensing:

Proprietary.

Contact:

calin@vulturelabs.com

The DNS has long relied upon serial number arithmetic, a concept which has never really been defined, certainly not in an IETF document, though which has been widely understood. This memo supplies the missing definition. It is intended to update RFC1034 and RFC1035. [STANDARDS-TRACK] This document defines algorithms for Authenticated Encryption with Associated Data (AEAD), and defines a uniform interface and a registry for such algorithms. The interface and registry can be used as an application-independent set of cryptoalgorithm suites. This approach provides advantages in efficiency and security, and promotes the reuse of crypto implementations. [STANDARDS-TRACK] This document defines TCP's four intertwined congestion control algorithms: slow start, congestion avoidance, fast retransmit, and fast recovery. In addition, the document specifies how TCP should begin transmission after a relatively long idle period, as well as discussing various acknowledgment generation methods. This document obsoletes RFC 2581. [STANDARDS-TRACK] This document defines the standard algorithm that Transmission Control Protocol (TCP) senders are required to use to compute and manage their retransmission timer. It expands on the discussion in Section 4.2.3.1 of RFC 1122 and upgrades the requirement of supporting the algorithm from a SHOULD to a MUST. This document obsoletes RFC 2988. [STANDARDS-TRACK] This document proposes an experiment to increase the permitted TCP initial window (IW) from between 2 and 4 segments, as specified in RFC 3390, to 10 segments with a fallback to the existing recommendation when performance issues are detected. It discusses the motivation behind the increase, the advantages and disadvantages of the higher initial window, and presents results from several large-scale experiments showing that the higher initial window improves the overall performance of many web services without resulting in a congestion collapse. The document closes with a discussion of usage and deployment for further experimental purposes recommended by the IETF TCP Maintenance and Minor Extensions (TCPM) working group. This memo specifies two elliptic curves over prime fields that offer a high level of practical security in cryptographic applications, including Transport Layer Security (TLS). These curves are intended to operate at the ~128-bit and ~224-bit security level, respectively, and are generated deterministically based on a list of required properties. This document describes elliptic curve signature scheme Edwards-curve Digital Signature Algorithm (EdDSA). The algorithm is instantiated with recommended parameters for the edwards25519 and edwards448 curves. An example implementation and test vectors are provided. Session Traversal Utilities for NAT (STUN) is a protocol that serves as a tool for other protocols in dealing with NAT traversal. It can be used by an endpoint to determine the IP address and port allocated to it by a NAT. It can also be used to check connectivity between two endpoints and as a keep-alive protocol to maintain NAT bindings. STUN works with many existing NATs and does not require any special behavior from them. STUN is not a NAT traversal solution by itself. Rather, it is a tool to be used in the context of a NAT traversal solution. This document obsoletes RFC 5389. In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document proposes a small modification to the way TCP increases its congestion window. Rather than the traditional method of increasing the congestion window by a constant amount for each arriving acknowledgment, the document suggests basing the increase on the number of previously unacknowledged bytes each ACK covers. This change improves the performance of TCP, as well as closes a security hole TCP receivers can use to induce the sender into increasing the sending rate too rapidly. This memo defines an Experimental Protocol for the Internet community. This document describes Virtual eXtensible Local Area Network (VXLAN), which is used to address the need for overlay networks within virtualized data centers accommodating multiple tenants. The scheme and the related protocols can be used in networks for cloud service providers and enterprise data centers. This memo documents the deployed VXLAN protocol for the benefit of the Internet community. This document describes a simple process that allows authors of Internet-Drafts to record the status of known implementations by including an Implementation Status section. This will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. This process is not mandatory. Authors of Internet-Drafts are encouraged to consider using the process for their documents, and working groups are invited to think about applying the process to all of their protocol specifications. This document obsoletes RFC 6982, advancing it to a Best Current Practice. Network virtualization involves the cooperation of devices with a wide variety of capabilities such as software and hardware tunnel endpoints, transit fabrics, and centralized control clusters. As a result of their role in tying together different elements of the system, the requirements on tunnels are influenced by all of these components. Therefore, flexibility is the most important aspect of a tunneling protocol if it is to keep pace with the evolution of technology. This document describes Geneve, an encapsulation protocol designed to recognize and accommodate these changing capabilities and needs. This document defines the core of the QUIC transport protocol. QUIC provides applications with flow-controlled streams for structured communication, low-latency connection establishment, and network path migration. QUIC includes security measures that ensure confidentiality, integrity, and availability in a range of deployment circumstances. Accompanying documents describe the integration of TLS for key negotiation, loss detection, and an exemplary congestion control algorithm. This document describes the data plane protocol for the Locator/ID Separation Protocol (LISP). LISP defines two namespaces: Endpoint Identifiers (EIDs), which identify end hosts; and Routing Locators (RLOCs), which identify network attachment points. With this, LISP effectively separates control from data and allows routers to create overlay networks. LISP-capable routers exchange encapsulated packets according to EID-to-RLOC mappings stored in a local Map-Cache. LISP requires no change to either host protocol stacks or underlay routers and offers Traffic Engineering (TE), multihoming, and mobility, among other features. This document obsoletes RFC 6830.

Acknowledgments The author thanks the participants of the IETF AI protocols discussions for their contributions to the understanding of the agent communication landscape.

Wire Examples

SYN Packet A SYN packet from 0:0000.0000.0001 port 49152 to 0:0000.0000.0002 port 1000, with no payload:

Total: 34 bytes.

Data Packet An ACK data packet with 5-byte payload "hello":

Total: 39 bytes.

Encrypted Tunnel Frame A PILS frame carrying an encrypted Pilot packet: