Internet anonymity without Tor

Lesswrong disclaimer

I wrote this before I was working full-time on ASI risk, and may not endorse people working on this stuff today. I currently lean a lot more in favour of making the world transparent, not private

2025-06-18

Disclaimer

As of 2025, there is no empirical evidence of successful deanonymisation attack on Tor. If any govt has this capability, they're successfully keeping this info private. There is public evidence of many of the internet's fiber optic cables being tapped.

Summary

Governments can wiretap fiber optic cables and obtain connections between senders and receivers, along with timestamps.
If senders send their pgp-encrypted messages to everyone, and the receiver retrieves the entire dump of all messages from one of these users same hours or days later, then this metadata is much harder to collect.
This setup is expensive hence it only works for <1 MB text payloads sent on >1 gbps connections.

Intelligence-agency-resistant internet anonymity is hard because the physical infrastructure can be inspected by someone with a monopoly on violence.

Fiber optic cables cannot hide sender/receiver identities as the attacker can wiretap the cables and then follow the physical path to identify which cable exactly carries a given message. Then they can break into the building that the cable enters.
- (also fiber optic connections usually requires KYC in most countries, but that's a legal limit not a physical one)
Radio signals cannot hide sender/receiver identities as the attacker can triangulate the signal based on signal strengths. Then they can break into the building that is transmitting the signal.
- (also encrypted radio is illegal in many countries, but that's a legal limit not a physical one)

Success criteria of attacker

When considering intelligence-agency-resistant anonymity, getting the metadata alone is enough to count as an attack, even if they don't get the message content.
Metadata includes sender/receiver irl identities, sender/receiver pseudonyms, message sizes and timestamps.
If the receiver is marked as suspicious, then any sender that connects with them is also marked as suspicious.

Attack 1: Get view access into majority of exit nodes

Tor relies on the sender passing each message to three other random users before it reaches the receiver, and hoping the three intermediaries don't all collude with the attacker.
If an intelligence agency has view access into majority of exit nodes, they can deanomyise Tor completely.
This could be done by controlling exit nodes themselves or by breaking into exit nodes run by others. They can do the latter using hardware or software backdoors, using targeted cyberattacks or using spies.

Attack 2: Wiretap source and receiver machine

If the intelligence agency is tapping fiber optic cables of both source and receiver, the timestamps of packets sent will match. This means they are aware of the physical addresses of both machines, the fact that there's a connection between them and the time interval in which this connection occured.

What if the sender just sent the message to everyone instead of sending it to their intended receiver?

Assume that some receivers may be publishing public proofs (via youtube, twitter etc) of their latest uncompromised PGP keys.
Assume that each user sends a single payload of X bytes to all users each day. This payload can include encrypted messages to specific users. If they have less than X bytes to send, they fill the remaining bytes with junk data.
Assume each user sends their X bytes at approximately the same time each day.
Only the actual receivers of the content can decrypt the message. It is junk to everyone else.
Assume 'gpg --hidden recipient' was used, so there's no way to tell which pubkey was used to encrypt a given message, from a given set of pubkeys.

Throughput

8 billion users, each user has 1 gbps unmetered fiber optic
- 1 gbps / 8B = 0.016 bytes/s = 1350 bytes/day
100 million users, each user has 1 gbps unmetered fiber optic
- 1 gbps / 100M = 105.4 KB/day
100 million users, each user has 10 gbps unmetered fiber optic
- 10 gbps / 100M = ~1.03 MB/day

Potential problems

Real-time messaging not possible. This is slow like courier.
Running servers from residential area requires effort. ISPs and OS developers can make this difficult. Renting a cloud server to download the messages does not work, as the cloud server owner knows which subset of these messages you downloaded to your local machine or display.