DDoS attacks on Tor: what is confirmed and why bridges suffer
DDoS attacks on Tor: what the Tor Project recorded since 2022, which defense shipped, what is still a theory, and why bridges do not speed up the network.
Tor has run under DDoS attacks more than once in recent years, and users noticed it themselves: Tor Browser took ages to connect and .onion sites loaded one try in two. The best documented episode began in 2022. The Tor Project kept an incident entry on its status page status.torproject.org. In August 2023 it released Tor 0.4.8 with a new defense for onion services, described on the Tor Project blog in the post “Introducing Proof-of-Work Defense for Onion Services”. Bridges, which people use to reach Tor from censored countries, by design hide only the entrance to the network, so they do not help when the network itself is overloaded. Below, confirmed facts are separated from theories, and theories are labeled.
How a DDoS on Tor differs from blocking Tor
The two problems are often confused, though they break Tor in different ways.
| Blocking | DDoS | |
|---|---|---|
| Who acts | A censor at the level of a country’s ISPs | Anyone with a botnet or many servers |
| What happens | The network cannot be reached | The network is reachable but clogged with junk traffic |
| Who is affected | People in one country | All Tor users worldwide |
| Remedy | Bridges: obfs4, Snowflake, WebTunnel | Defenses inside the network and on sites |
Bridges were built against blocking. They were never built against DDoS. This article is about DDoS. How Russia, Iran and China block Tor is a separate topic, and mixing the two makes it easy to misread the cause: a slow Tor in Tehran can be a block or a problem for the whole network.
Why Tor can be overloaded at all
Tor runs on volunteers. Individuals, universities and nonprofits operate relays and pass other people’s traffic through them. There are several thousand relays (estimate: 7,000-8,000) and fewer bridges (estimate: 2,000-3,000). Tor Metrics publishes exact daily numbers. The Tor Project owns no data centers.
A user’s connection passes through three relays. The first sees who connected but not where they are going. The last sees where the traffic goes but not who sent it. The middle sees neither. That is what anonymity rests on.
The flip side: there is no central point in the network where a shared filter can be installed, the way Cloudflare does for ordinary websites. Each server defends itself, and nobody sees the attack as a whole.
The core weakness is uneven cost. Asking a relay to build a new route costs almost nothing, while the relay spends noticeable computing work on each request. Think of a restaurant where tables can be booked by phone for free: one call means nothing, a thousand robot calls a minute leave real guests without a seat. The attacker does not need to break any encryption. It is enough to make the network work for nothing.
Why onion services suffer more
A site with a .onion address has no public IP. Connecting to it involves several extra steps, and part of the work is done by the site itself. Every incoming request costs it resources. When a single .onion site gets a flood of such requests, it stops answering everyone at once. That is why onion services became the main victim of the 2022-2023 attacks.
What the Tor Project confirmed about the 2022-2023 attacks
| When | What | Where to check |
|---|---|---|
| 2022, starting in summer (estimate based on the status page entry) | Prolonged network performance problems caused by DDoS | status.torproject.org, incident history |
| 2022-2023 | Onion services were hit hardest | Tor Project blog |
| August 2023 | Tor 0.4.8 ships proof-of-work protection for onion services | Tor Project blog, post “Introducing Proof-of-Work Defense for Onion Services” |
What the checked Tor Project publications do not contain:
- the size of the attack in gigabits per second or number of requests;
- the size of the botnet and the countries the traffic came from;
- the attacker’s name and motive;
- an exact end date. Based on the status page entries, the episode lasted about a year, which is an estimate.
The missing numbers say something too. Commercial DDoS protection services love publishing terabit records because they see all of a customer’s traffic. The Tor Project cannot: each relay knows only its own share of the load, and the network’s design does not allow adding it up into one figure.
How the proof-of-work defense works
This is the Tor Project’s main answer to the attacks, and the idea is simple.
When a .onion site is under load, it asks every visitor to solve a small computing puzzle before connecting. Tor Browser does it automatically, and the person sees only a short delay. The heavier the load, the harder the puzzle. Visitors who put more effort into solving it are served first.
For a person, that is one puzzle per connection. For a bot sending millions of requests, it is millions of puzzles, which means real spending on processors and electricity. The attack does not become impossible. It becomes expensive. It is the same logic as a captcha, minus the pictures of traffic lights.
Limits of the defense:
- it works only for .onion sites, ordinary sites opened through Tor gain nothing from it;
- the onion site’s owner has to turn it on in the settings;
- it does not protect the relays that carry all of the network’s traffic.
Why a bridge does not save you from DDoS
A bridge solves one problem: hiding from the censor that a person is connecting to Tor. This follows from the design described in Tor Project documentation.
A route through a bridge looks like this: user, bridge, middle relay, then an exit relay or an onion site. The bridge takes the place of the first relay. Everything after it is shared by all users. If middle relays or the .onion site itself are overloaded, traffic through a bridge waits in the same queue as traffic without one.
As far as could be checked, the Tor Project has not published separate measurements of bridge speed during DDoS. So the conclusion above is logical, not measured.
| Transport | How it disguises traffic | What it hides from the censor | Helps when the network is overloaded |
|---|---|---|---|
| obfs4 | Traffic looks like random noise | The fact of connecting to Tor | No |
| Snowflake | Looks like a video call, proxies keep changing | The entry address, which is always new | No |
| WebTunnel | The bridge looks like an ordinary HTTPS site | That the server is linked to Tor | No |
| meek | Traffic goes through a large CDN | The bridge’s real address | No |
The choice of bridge decides whether the censor lets you through. Speed during an attack on the network barely depends on it.
Theories: where bridges may be more vulnerable than relays
What follows is reasoning, not confirmed fact. No public cases testing these ideas in practice could be found in Tor Project materials.
Theory 1. Bridges run on weaker hardware. Many bridges are run by volunteers on cheap VPS. If that holds broadly, such servers have less headroom than large relays on university networks.
Theory 2. Bridge addresses can be collected. Bridges are handed out through bridges.torproject.org, email and a Telegram bot. Censorship researchers have described censors trying to collect addresses through these channels in order to block them. In theory, the same list works for an attack.
Theory 3. The hoster decides a bridge’s fate. To a hoster, a flood aimed at a bridge looks like an attack on a customer’s server. A common hoster response to heavy flooding is to null-route the IP for a while. How many bridges dropped out this way during attacks is not publicly known.
Theory 4. Snowflake has a broker. Snowflake’s temporary proxies, which are volunteers’ browsers, are matched with users by a separate broker server. That much is documented. That the broker could become a target for overload is an inference from the design, not a recorded incident.
Who is behind the attacks on Tor: theories
The Tor Project publications that could be checked do not name a culprit, and nobody has published verifiable evidence. Everything below is theory.
- Darknet market rivalry. In 2022-2023 onion services took the main hit. That fits the theory of marketplaces fighting each other and extortion. The Tor Project did not confirm it.
- States. A slow Tor suits censoring countries even without a full block: a person whose page takes a minute to load gives up. There is no direct evidence that Russia, Iran, China or any other state took part in DDoS against Tor.
- Deanonymization. With some relays knocked out, clients build new routes more often, which raises the chance of landing on an attacker’s relay. Tor counters this with persistent entry relays that a client keeps using for a long time. No confirmed public cases of this scheme exist.
- Botnet testing. Tor is a large, always-on test bed. A botnet that can slow it down can be sold afterwards. Also just a theory.
The 2024 episode: what still needs checking
In late 2024, Tor relay operators publicly reported a wave of hoster complaints about suspicious activity coming from relay addresses. According to these reports, relay addresses were forged to turn hosters against them. If confirmed, this is not a classic DDoS but an attack through the hosting abuse desk. This article does not give exact dates, scale or the Tor Project’s official assessment: they need to be checked against the primary source.
Where to check Tor’s status yourself
- status.torproject.org: current incidents and their history, including DDoS entries.
- Tor Metrics: users by country and by bridge type. A drop in every country at once looks like a network problem, a drop in one country looks like a block.
- OONI Explorer: volunteer-collected tests of Tor and Snowflake reachability by country.
What is confirmed and what is not
Confirmed:
- since 2022 the Tor Project has publicly recorded DDoS and network performance problems on its status page;
- onion services were hit hardest;
- in August 2023 Tor 0.4.8 shipped with proof-of-work protection for onion services;
- by design, a bridge replaces only the entrance to the network.
Not confirmed:
- who ordered the attacks and why;
- state involvement;
- links to specific darknet marketplaces;
- use of DDoS for deanonymization;
- how much more bridges suffer than ordinary relays;
- the scale and details of the 2024 episode.