Files
ngit-grasp/docs/explanation/defensive-measures.md
T
DanConwayDev 6a3b723eaf fix(sync): restrict event-directed sync targets to globally reachable endpoints
After deploying a8964bb to gitnostr.com, production logs showed
event-directed proactive sync dialling ws://localhost:3334,
ws://127.0.0.1:7334, and ws://100.125.184.46:7334 (CGNAT). Repository
announcements, state events, and PR events are untrusted - anyone can
publish them - yet their relays/clone tags reached the outbound
WebSocket and git-fetch sinks after syntax-only checks, letting a
crafted event point a public relay at loopback, private, link-local,
or local-name infrastructure (SSRF).

Add one fail-closed outbound target policy (src/outbound.rs) applied
immediately before every event-directed sink so no call path can
bypass it:

- RelayConnection::connect re-authorizes (with DNS vetting) before
  every dial and reconnect. SyncManager::register_relay additionally
  refuses to register forbidden targets so they never enter the
  reconnect lifecycle, and memoizes rejections so stored events cannot
  spam logs or starve the bounded purgatory sync tick.
- RealSyncContext::fetch_oids authorizes clone URLs from announcements
  and purgatory PR events immediately before spawning git fetch, then
  pins the vetted DNS answers via http.curloptResolve and confines the
  subprocess with GIT_ALLOW_PROTOCOL=http:https,
  http.followRedirects=false, cleared proxy config/environment, and an
  empty credential helper, so redirects, proxies, or alternate
  protocols cannot escape the authorized target.

The policy enforces per-sink scheme allowlists (ws/wss for relays,
http/https for git), rejects embedded credentials and local hostnames
(localhost, single-label names, IANA special-use suffixes), and
requires IP literals and every DNS answer to be globally reachable.
Service admission (lists_service) and the don't-fetch-from-ourselves
filter now compare parsed host and port instead of substrings, so
gitnostr.com.attacker.example or a path containing the domain no
longer satisfies a check for gitnostr.com.

The operator-configured bootstrap relay stays usable even when local:
trust is carried by RelayTargetSource::OperatorConfigured at
construction, never by comparing event URLs against the configured
value, so event URLs that merely resemble the bootstrap relay are
still rejected. The new NGIT_SYNC_ALLOW_NON_GLOBAL_TARGETS option
(default false; documented in configuration.md, module.nix, and
.env.example) relaxes only the reachability checks for integration
tests and closed development networks; the TestRelay fixture sets it
because the test infrastructure lives on loopback, while the new
regression tests opt back into production behaviour.

Known limitation: nostr-sdk's connect API takes a URL, not a
pre-resolved address, so relay DNS is re-validated before every dial
but re-resolved by the SDK during connection, leaving a narrow
DNS-rebinding window (documented in defensive-measures.md). Git
fetches do not share this window because their DNS answers are pinned.
Closing it requires upstream connector support rather than a custom
connector here.

Validation: tests/outbound_policy.rs adds integration scenarios
against the real relay binary proving that loopback relay URLs and
loopback git clone URLs produce no outbound connection (counting TCP
listeners stand in for attacker infrastructure), that private,
link-local, CGNAT, unspecified, and multicast literals plus localhost
and credential URLs are rejected, that the local bootstrap relay still
connects while a resembling event URL is rejected, and that
substring-embedded domains are no longer admitted. src/outbound.rs
unit tests cover the reachability matrix (including 100.125.184.46)
and exact service matching. cargo fmt, cargo clippy (workspace, zero
warnings), the full cargo test suite, and cargo test -p grasp-audit
--lib all pass in the nix dev shell.
2026-08-01 19:39:36 +00:00

10 KiB

Defensive Measures & Rate Limiting

This document describes the defensive measures implemented in ngit-grasp to protect against abuse, spam, and denial-of-service attacks.

Note: A point-in-time analysis of defensive measures in other Nostr relays (strfry, nostr-rs-relay, khatru) was conducted to inform these design decisions. The analysis examined connection limits, rate limiting approaches, and per-IP enforcement strategies across the ecosystem.

Overview

ngit-grasp employs multiple layers of defense:

  1. Connection & Subscription Limits - Per-connection limits on subscriptions and event publishing
  2. Content Filtering - Blacklist/whitelist system for repositories and event authors
  3. Event Validation - Strict GRASP-01 protocol validation
  4. Relay Health Management - Intelligent handling of problematic remote relays

What's Implemented

Per-Connection Rate Limits

Source: Built-in to rust-nostr relay-builder

  • Subscription limit: Max 500 concurrent subscriptions per connection
  • Event publishing limit: Max 60 events per minute per connection
  • Subscription ID length: Max 250 characters
  • Filter limit: Max 500 results per query (default)

These limits prevent individual connections from overwhelming the relay.

Per-IP Connection Monitoring

Source: Custom ngit-grasp implementation
Location: src/metrics/connection.rs

  • Status: Monitoring only (does NOT enforce limits)
  • Tracks connections per IP address internally
  • Flags IPs exceeding threshold (default: 10 connections)
  • Privacy: IP addresses never exposed in Prometheus metrics, only aggregate counts
  • Logs warnings when threshold exceeded

Note on enforcement: Per-IP connection limits are not built into rust-nostr relay-builder (tracks per WebSocket connection, not per IP). If abuse is detected via metrics, enforcement should be implemented as a PR to rust-nostr/relay-builder to benefit the entire Nostr ecosystem, rather than custom code in ngit-grasp.

Content Filtering (Blacklists/Whitelists)

Source: Custom ngit-grasp implementation
Location: src/config.rs, src/nostr/builder.rs

Event Blacklist:

  • Block ALL events from specific authors (npubs)
  • Takes precedence over all other validation
  • Events never reach storage or purgatory

Repository Blacklist:

  • Block specific repositories, developers, or identifiers
  • Takes precedence over whitelists
  • Three formats: npub, npub/identifier, identifier

Repository Whitelist:

  • Curate which repositories are accepted (GRASP-01 mode)
  • Only accept announcements that both list your service AND match whitelist
  • Same three formats as blacklist

Archive Whitelist (GRASP-05):

  • Mirror specific repositories even if they don't list your service
  • Same three formats as blacklist
  • Default: read-only mode when enabled

Privacy: Blacklists not advertised in NIP-11 metadata.

Event Validation Plugin System

Source: Built-in to rust-nostr relay-builder
Implementation: Custom GRASP-01 validation in src/nostr/builder.rs

  • WritePolicy trait: Controls which events are accepted
  • QueryPolicy trait: Controls which queries are allowed (not currently used)
  • Access to client IP address for future per-IP rate limiting
  • Modular sub-policies for different event types (announcements, state events, PRs)

Relay Health Management (GRASP-02 Sync)

Source: Custom ngit-grasp implementation
Location: src/sync/health.rs

Exponential Backoff:

  • Failed connections trigger increasing delays: 5s → 10s → 20s → ... → 1 hour max
  • Prevents hammering dead or slow relays

Naughty List:

  • Tracks relays with persistent infrastructure issues (DNS, TLS, protocol errors)
  • Separate from normal connection failures
  • 12-hour expiration (configurable)
  • Reduces retry frequency for broken relays

Rate Limit Detection:

  • Detects when remote relay rate limits us
  • Automatic 65-second cooldown
  • Prevents hammering relays that tell us to slow down

Domain Throttling (Git Data Fetching):

  • Max 5 concurrent requests per domain
  • Max 30 requests per minute per domain
  • Respectful rate limiting when fetching missing git data

Outbound Target Policy (SSRF Protection)

Source: Custom ngit-grasp implementation
Location: src/outbound.rs

Repository announcements, state events, and PR events are untrusted input, but their relays and clone tags direct proactive sync's outbound WebSocket connections and purgatory git fetches. Production logs showed event-directed sync dialling ws://localhost:3334, ws://127.0.0.1:7334, and ws://100.125.184.46:7334 (CGNAT); a crafted event could use this to probe or attack the relay host's local network.

One fail-closed policy is applied immediately before every event-directed outbound sink:

  • Relay connections (RelayConnection::connect): re-checked before every dial, including reconnects, plus a registration-time gate so forbidden targets never enter the reconnect lifecycle.
  • Git fetches (RealSyncContext::fetch_oids): checked immediately before spawning git fetch.

The policy enforces:

  • Scheme allowlist per sink - ws/wss for relays, http/https for git; no file:, ssh:, git:, or other protocol escapes.
  • No credentials in URLs.
  • No local hostnames - localhost, single-label names, and IANA special-use suffixes (.local, .internal, .home.arpa, .onion, .test, .invalid, .localdomain).
  • Globally reachable addresses only - loopback, private (RFC 1918), CGNAT (RFC 6598), link-local, unspecified, multicast, broadcast, documentation, benchmarking, reserved, and unique-local ranges are all rejected, for IP literals and for every DNS answer (resolution failure fails closed).
  • Exact service matching - GRASP-01 admission (lists_service) and the don't-fetch-from-ourselves filter compare parsed host and port, so gitnostr.com.attacker.example or https://evil.example/gitnostr.com/ cannot satisfy a check for gitnostr.com.

The operator-configured bootstrap relay (NGIT_SYNC_BOOTSTRAP_RELAY_URL) is trusted and exempt, so a local bootstrap relay keeps working. Trust is carried by the source of the URL, never by URL comparison: an event-provided URL that merely resembles the bootstrap relay is still rejected.

Git subprocesses are additionally confined so an authorized URL cannot escape the decision: GIT_ALLOW_PROTOCOL=http:https, http.followRedirects=false, proxies disabled (config and environment), credential.helper= cleared, and the vetted DNS answers pinned via http.curloptResolve.

Known limitation (DNS rebinding, relay connections only): the vetted DNS answers cannot be pinned onto nostr-sdk's connector because its connect API accepts a URL, not a pre-resolved socket address. Relay DNS is therefore re-validated immediately before every connection attempt but re-resolved by the SDK during the dial, leaving a narrow time-of-check/time-of-use window. Closing it needs upstream connector support for pre-resolved addresses (rust-nostr). Git fetches do not share this window because their DNS answers are pinned.

The escape hatch NGIT_SYNC_ALLOW_NON_GLOBAL_TARGETS=true disables the reachability checks (scheme and credential checks remain) for integration tests and closed development networks. Production relays must leave it unset.

What's NOT Implemented

Per-IP Rate Limiting

  • Per-IP connection limits: Not enforced (only monitored)
  • Per-IP subscription limits: Not supported
  • Per-IP event publishing limits: Not supported

Why: rust-nostr relay-builder tracks limits per WebSocket connection, not per IP address.

To implement: Would require custom middleware/WritePolicy to aggregate across connections from the same IP.

Query Filtering

Status: QueryPolicy trait available but not currently used.

Potential uses: Rate limit queries per IP, block expensive queries, restrict access to certain event kinds.

Future Enhancements

Per-IP Rate Limiting

Per-IP connection and event rate limiting were considered but deferred until abuse is detected in production. The current protections (per-connection limits, total connection limit, content filtering) are sufficient for the git relay use case.

Decision rationale: The primary DoS vector is connection exhaustion, which is addressed by the total connection limit (NGIT_MAX_CONNECTIONS). Per-IP enforcement would require custom middleware in rust-nostr relay-builder (which currently tracks limits per WebSocket connection, not per IP). If abuse is detected via the per-IP monitoring metrics, enforcement should be implemented as a PR to rust-nostr/relay-builder to benefit the entire Nostr ecosystem.

Related: Git endpoint throttling (issue ff38) is a separate concern with different requirements.

Summary Table

Feature Status Enforced? Configurable?
Per-Connection Limits
Max subscriptions (500) ✅ Active Yes No (relay-builder default)
Event rate limit (60/min) ✅ Active Yes No (relay-builder default)
Total Connection Limit
Max connections (500) ✅ Active Yes Yes (NGIT_MAX_CONNECTIONS)
Per-IP Monitoring
Connection tracking ✅ Active No (monitor only) Threshold only
Content Filtering
Event blacklist ✅ Active Yes Yes
Repository blacklist ✅ Active Yes Yes
Repository whitelist ✅ Active Yes (if set) Yes
Archive whitelist ✅ Active Yes (if set) Yes
Event Validation
GRASP-01 validation ✅ Active Yes Via WritePolicy
Relay Sync Protection
Exponential backoff ✅ Active Yes Yes
Naughty list ✅ Active Yes Yes (12h default)
Rate limit detection ✅ Active Yes Automatic
Domain throttling ✅ Active Yes Hardcoded (5/30)
Outbound target policy (SSRF) ✅ Active Yes Yes (NGIT_SYNC_ALLOW_NON_GLOBAL_TARGETS)
Not Implemented
Per-IP connection limit ⚠️ Deferred No -
Per-IP rate limiting ⚠️ Deferred No -
Query filtering ⚠️ Available No Not implemented