Skip to content

Enterprise agent platform for building, deploying, and operating AI teammates.

3.9 KB
Raw Blame

Authentication rate limits and deployment

Before starting a production/staging API, configure a reachable XAGENT_REDIS_URL on every replica. Startup pings Redis and refuses admission if it is absent, malformed or unreachable. Runtime failures fail closed with a sanitized HTTP 503. A Redis command has a two-second connection/read timeout and no automatic retry/backoff. Only development without secure-secret enforcement can use bounded in-process budgets, including during an outage. Local fallback does not provide cross-process protection.

Budgets use a fifteen-minute sliding window:

Endpoint Source IP Resolved account/email IP + account/email
Login 200 50 5
Forgot password 200, shared with reset 50 5
Reset password same shared recovery budget — —

Username and email login aliases share the resolved user-ID budget. Recovery email budgets are deliberately higher than the browser/email-pair budget so five unauthenticated requests cannot globally lock an email. Reset guesses all share the source-IP budget; arbitrary token strings never allocate buckets. Short reset passwords are rejected before consuming rate-limit capacity.

Login first reserves source-IP capacity before opening an auth database session. Only admitted requests resolve the account and atomically reserve account/pair budgets. The pre-admission budget counts failures and in-flight work, including requests subsequently rejected by the identity budgets. On successful login, only that request's IP reservation is removed; other users' failures remain. This prevents many successful logins behind corporate NAT or an unknown peer from cumulatively exhausting the IP budget. Refund failure never fails an already-committed successful login.

Recovery reserves its shared IP/email/pair budgets atomically and counts all requests, including generic successful forgot-password responses: refunding those would disable email-abuse protection and reveal account existence. Forgot/reset intentionally share one IP budget. Rejection returns HTTP 429 with Retry-After equal to the maximum remaining wait across exhausted buckets in that reservation.

Configure positive integer limits in config.py through:

  • XAGENT_AUTH_RATE_LIMIT_WINDOW_SECONDS (900)
  • XAGENT_AUTH_LOGIN_IP_LIMIT (200), XAGENT_AUTH_LOGIN_ACCOUNT_LIMIT (50), XAGENT_AUTH_LOGIN_PAIR_LIMIT (5)
  • XAGENT_AUTH_RECOVERY_IP_LIMIT (200), XAGENT_AUTH_RECOVERY_EMAIL_LIMIT (50), XAGENT_AUTH_RECOVERY_PAIR_LIMIT (5)
  • XAGENT_AUTH_RATE_LIMIT_MAX_LOCAL_KEYS (10000, development only)

Use consistent configuration across replicas. Size limits for actual NAT/proxy traffic and configure XAGENT_TRUSTED_PROXY_HOPS only for trusted proxies; do not blindly trust forwarded addresses to work around an incorrectly identified peer. Invalid/non-positive values fall back to defaults, never disabling protection.

Accepted trade-off: resolved user-ID account budgets prevent username/email alias bypass, but saturating a known account can reveal an alias relationship through 429 versus 401. These generic response bodies are not a claim of complete account enumeration resistance; monitor distributed abuse and assess this privacy trade-off for the deployment. Switching to identifier-only keys would reopen alias bypass.

Redis keys expire independently; there is no global active-identity cap that one source can exhaust to reject unrelated users. Plan Redis memory/availability and use an appropriate dedicated namespace/service; do not use an eviction policy that silently drops active security budgets. The plain Redis client targets standalone Redis or a managed proxy, not RedisCluster discovery. Local development evicts expired or unlocked buckets at capacity, never active lockouts. If every bounded slot is locked it returns a distinct 503 capacity error rather than unlocking an attacker or claiming an identity is rate-limited.