ATProto has an open relay problem. The trajectory is following email's history with uncomfortable precision, and the latest escalation -- scammers deploying lookalike PDS domains to get innocent operators relay-banned -- shows the pattern accelerating.
The parallel
In the early days of email, any mail server could relay messages for any sender. This was a feature: it made email easy to deploy and federate. It was also an exploit: spammers sent bulk mail through open relays, turning other people's servers into spam infrastructure.
The response was blacklists. Organizations like Spamhaus maintained lists of IP addresses known to relay spam. Mail servers checked incoming connections against these lists and rejected traffic from blacklisted IPs. This worked for stopping spam but created collateral damage: innocent servers sharing IP ranges with spammers got blacklisted. Running your own mail server became a reputational liability.
ATProto PDS hosting is following the same trajectory:
- Open creation. Most PDSes allow account creation with minimal verification. This is a feature -- low barriers to entry are what make the network accessible.
- Spam exploitation. Spammers discovered they could mass-create accounts on unmonitored PDSes, using them as infrastructure for bulk posting, follow-spam, and phishing.
- Domain-level bans. Relay operators responded by banning entire PDS domains. If a PDS is producing spam accounts, the relay stops ingesting its records.
- Lookalike attacks. The latest escalation: scammers now deploy PDSes with domains that mimic legitimate operators. A fake version of a well-known community PDS creates spam accounts, gets the lookalike domain banned, and potentially damages the reputation of the legitimate operator.
- Collateral damage. Innocent PDS operators on the same hosting infrastructure, or with similar domain names, risk being caught in relay-level bans they did not cause and cannot easily contest.
Why this matters for decentralization
The open relay parallel has a specific endpoint in email history: most people stopped running their own mail servers. The reputational cost became too high. Today, the vast majority of email flows through a handful of providers (Gmail, Outlook, Yahoo) who have the resources to maintain sender reputation, implement DKIM/SPF/DMARC, and negotiate with blacklist operators.
If ATProto follows the same path, PDS decentralization contracts. Running your own PDS already requires technical skill. If it also requires maintaining a spam-free reputation with relay operators -- and defending that reputation against lookalike attacks -- the barrier rises further. The equilibrium becomes a small number of large PDS operators who can afford the operational overhead, plus a long tail of self-hosters who accept the risk.
This is exactly the dynamic the Shannon Entropy dashboard (arewedecentralizedyet.online) would capture: PDS count might remain high while account concentration increases, because small operators quit under reputational pressure.
What email's history suggests
Email solved the open relay problem through a combination of mechanisms:
- Authentication protocols (SPF, DKIM, DMARC) -- cryptographic proof that a message came from an authorized sender for that domain. ATProto already has this: records are signed by the PDS operator's key, and DIDs provide verifiable identity.
- Reputation systems -- sender reputation scores based on historical behavior. This does NOT exist for ATProto PDSes. There is no shared, queryable reputation system that relay operators can consult before deciding whether to ingest a PDS's records.
- Rate limiting -- throttling new connections and message volume from unknown senders. ATProto relay operators can implement this per-PDS, but there is no standardized protocol for PDS rate limiting or graduated trust.
- Feedback loops -- ISPs reporting spam complaints back to senders so they can clean up. ATProto has no equivalent mechanism for relay operators to signal abuse concerns to PDS operators short of a full ban.
The authentication layer exists. The reputation, rate limiting, and feedback layers do not.
The missing infrastructure
Two pieces of infrastructure would address the gap:
PDS reputation signals. A shared, queryable system where relay operators, AppView operators, and moderation services can publish and consume reputation assessments about PDS domains. This could be built on ATProto itself -- reputation records as governance records, signed by the assessing party, aggregatable by any interested consumer. The governance record layer that already handles moderation labels and data reuse intents could extend to PDS reputation.
Graduated trust for new PDSes. Instead of binary allow/ban, relay operators could implement graduated trust: new PDSes start with limited throughput and earn higher limits based on account quality over time. This parallels email's "IP warming" practice where new mail servers gradually increase sending volume to build reputation.
Neither requires protocol changes. Both could be built as infrastructure services on top of the existing protocol. The pds-threat-feed proposal in various workspace tooling collections addresses part of this -- sharing abuse signals between PDS operators. But the relay-operator-facing reputation system is the missing piece.
The uncomfortable question
The open relay problem in email was solved. The cost was centralization. The question for ATProto is whether the same problem can be solved without the same cost -- whether PDS reputation systems can be decentralized enough that small operators can participate, or whether reputation management inevitably concentrates in the hands of the largest operators.
The governance record layer pattern suggests a possible answer: signed reputation records, published by multiple independent assessors, consumable by any relay operator. No single reputation authority. But this requires someone to build it, and it requires relay operators to adopt it. The current path -- ad hoc domain bans by a single primary relay operator -- is the centralization-by-default outcome that email already demonstrated.