Résumé
QTINeon has unauthenticated relay-to-host amplification via unbounded RECONNECT_REQUEST forwarding
Détails de l’avis
Impact
The relay's reconnect handler forwards every RECONNECT_REQUEST to the host without deduplication or a size cap on the pendingReconnects map, unlike the connect flow which guards against this with maxPendingConnections. An unauthenticated attacker who knows a valid session ID can send RECONNECT_REQUEST packets from many spoofed source addresses; each packet that passes the session lookup is forwarded to the host as a new reconnect attempt. Because the per-source rate limiter assigns a fresh token bucket to each spoofed IP, it provides no protection. The host receives one forwarded packet per spoofed source per cleanup cycle, making the relay an amplification vector for denial-of-service against the host. The host's real address is never exposed to clients by design, so this is the primary viable DoS path against it. A secondary issue: once more than maxRateLimiters spoofed IPs are seen, performCleanup calls rateLimiters.clear(), resetting rate limit state for all sources including legitimate ones.
Affected: NeonRelay in all three implementations (Java, Python, TypeScript).
Patches
Not yet patched. Fix consists of three changes to handleReconnectRequest in each implementation:
- Reject the request if
pendingReconnects.size() >= maxPendingConnections(mirrors the existing connect-flow guard) - Only forward to the host if the
sessionId:clientIdkey is not already present inpendingReconnects— the map entry can still be updated with the new source address, but the host only needs to validate once per slot - In
performCleanup, evict throttled entries before falling back torateLimiters.clear()to avoid resetting legitimate sources' rate limit state
Workarounds
Operators can partially mitigate by placing the relay behind a network-level filter that drops packets with spoofed source addresses (BCP38/uRPF). This does not address the missing pendingReconnects size cap or the rateLimiters.clear() issue but eliminates the amplification path in most deployment environments.
References
- PROTOCOL.md — reconnect flow
- ARCHITECTURE.md — reconnect handshake detail (step 4 describes the unbounded
put)
Références
Vulnérabilités liées
Tout Supply chain →- MEDIUMCVE-2026-71486
vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output bounds
- HIGHCVE-2026-67446
Mailpit: Thumbnail generation decodes unbounded image dimensions before scaling
- MEDIUMCVE-2026-73228
Django REST framework: Potential bypass of Django `DATA_UPLOAD_MAX_MEMORY_SIZE` when parsing oversized JSON and urlencoded request bodies via DRF `request.data`
- MEDIUMCVE-2026-55407
Buffa Vulnerable to Memory Exhaustion Denial of Service in decode_unknown_field via Unbounded Allocation
- MEDIUMCVE-2026-55531
PraisonAI MCP HTTP server has unauthenticated unbounded session accumulation (memory exhaustion; session TTL never enforced)
- HIGHCVE-2026-61827
netty-incubator-codec-ohttp: BinaryHttpParser should enforce limits for variable lengths fields