Tor Relay Attacks & Potential Deanonymization
Timeline, technical observations, relevant research, and a cautious threat assessment
This report synthesizes public relay-operator posts, Tor Project announcements, and research papers. It assesses whether the September–November 2026 attack activity may be more than availability disruption and could intersect with traffic-analysis or deanonymization techniques.
Executive summary
Independent research demonstrates that Guard discovery, DoS-assisted onion-service deanonymization, and exploitation of public DoS-related signals are technically possible. This research makes the incident pattern more concerning, but capability and incident linkage must not be conflated.
Observed with relatively high confidence
- Adaptive, multi-vector relay attacks.
- Repeated impact on Guard/HSDir roles.
- Circuit creation and directory request surges.
- Some reported descriptor-fetch concentration.
- Relay instability and loss of network flags.
Not established publicly
- Attacker identity or shared command-and-control.
- Confirmed client-to-service correlation by the campaign.
- Progressive packet/payload-size probing.
- Confirmed victim deanonymization.
- Connection between exit-relay compromise and DoS campaigns.
Scope note: This snapshot is based on public material available through 18 November 2026. Operator measurements are self-reported and may have different configurations, counters, baselines, time zones, and collection methods. Some posts were edited to correct earlier interpretations; this report prioritizes later corrections where visible.
Chronology
Later anti-censorship discussions described anomalous Snowflake user-count behavior without a corresponding increase in bandwidth or client polling. A directory-request attack was suggested as one possible explanation. This is a possible precursor, not evidence that the same campaign was active then.
Operators reported directory-fetch floods from relay sources, disproportionate traffic, and changing upload/download ratios. One report described ratios around 4:1–8:1 upload/download; after disabling directory caching (DirCache 0), the pattern reportedly flipped to about 1:2 download/upload and continued for 15–60 minutes before stopping. The operator speculated that trackable packet delays could be used to backtrace connections. That is an operator hypothesis, not demonstrated correlation.
Tor disclosed fixes including a stream-isolation issue: a malicious onion service or HSDir relay could potentially cause a circuit established under one isolation context to be reused by a later stream from another context. The release is relevant to the threat environment, but the vulnerability is not evidence of exploitation in the current attacks or a causal connection to them.
Reporting describes an Ebury malware-related compromise of relay-host systems and unauthorized exit-policy changes, redirecting SSH connections through interception servers. The number varies by counting scope: reports describe 333 relays in the interception campaign and broader counts of 377 relays with policy changes. Peak compromised exit weight was reported around 7.46%. This was a real host-compromise/interception incident; no public evidence links its operators to the DoS campaign.
Relay operators reported high-rate BEGIN_DIR floods after some relays acquired HSDir status. Reports include rates in the thousands per second, high circuit counts, CPU and memory pressure, reachability failures, and relays losing network flags. These observations suggest role-associated targeting, but targeted availability disruption remains a sufficient alternative explanation.
One operator documented attacks on two relays at different times, with reported HSDir loss at approximately 01:22:25 and 01:22:30 UTC—five seconds apart. Traffic ratios and BEGIN_DIR rates differed across the relays. This is suggestive of coordination, but consensus/monitoring timestamps are not necessarily the exact moment of flag change.
Additional operators reported attacks on Guard relays, recurring attack/recovery cycles, and consideration of shutting down groups of relays to limit damage. Some posts described attacks resuming after relays regained status. Role-targeting is a strong observation; deanonymization intent remains an inference.
Reports described BEGIN_DIR floods (including ~1,300/s, later ~2,100/s and higher), CPU saturation, memory growth, relay instability, and consensus exclusion. Tor released 0.4.9.15 on 28 October with BEGIN_DIR rate-limiting defenses and changes intended to make HSDir status more resilient. The defenses did not eliminate later attack vectors.
One operator reported ~4,400 BEGIN_DIR/s and another reported extreme circuit counts, memory growth, restarts, and high CPU. In one report, most inbound connections were from consensus relays, complicating defenses that focus on client sources.
A relay reported ntor_v3 handshake rates increasing from ~8/s to ~190–215/s, then decreasing to ~140–160/s and stepping up to ~230/s. Open circuits peaked around 22,000 versus a 3,000–5,000 baseline. Traffic was comparatively symmetric and modest; BEGIN_DIR counters stayed near baseline. Most CREATE cells reportedly came from consensus relays, and the step-up was attributed to existing sources increasing activity. The source later corrected earlier claims about bandwidth-authority measurement effects.
Operators reported a burst with tens of thousands of open circuits and a circuit-creation defense rejecting thousands of circuits in a window. Follow-up discussion distinguished relay-sourced circuit creation from client-sourced traffic and corrected interpretations of Guard/Measured changes. This illustrates why the thread's edits and operator corrections should be retained in any analysis.
Reports described ingress-heavy and egress-heavy attacks, followed by a traffic-collapse/unresponsive-relay pattern. In a later update, one operator reported 17.3 million hs_get requests from known relays, only 538 streams without a corresponding request, and more than 90% of descriptor fetches repeatedly fetching a small set of “hot” descriptors—estimated at roughly a dozen. Each fetch reportedly generated a fresh three-hop circuit. The activity faded after the relay lost HSDir. Another HSDir reportedly handled NTor v3 successfully during the same window, which is consistent with hash-ring/descriptor-specific concentration but does not prove target selection by service identity.
Important correction: a separate report's “exit-connection flood” interpretation was later qualified because BEGIN=exit-initiated and created≈closed can be normal; that operator lacked pre-attack baseline data. Avoid treating those exit counters alone as proof of a flood.
Research relevant to deanonymization
Guard discovery via HSDir — demonstrated capability
The 2022 paper From “Onion Not Found” to Guard Discovery demonstrated a web-based attack using requests for nonexistent onion addresses. Under the paper's evaluated conditions, an adversary with a small number of HSDirs and 5% of Tor relay bandwidth identified guards for 50% of victims in 12.06 seconds and 90% in 22.01 seconds. This is a proven technique, not evidence it is being used in the current campaign.
Descriptor Flood / Tordos — DoS as part of an anonymity attack
The 2026 IEEE S&P paper Descriptors of Exposure: Undermining Tor Anonymity Through Exploiting Descriptor Flood describes a descriptor-flooding and selective-node-teardown strategy. The authors report disabling nodes/services within minutes and exposing an onion service identity in 6.1 hours under their experimental conditions. This shows that selective DoS can be an integrated deanonymization method; it does not establish that the present campaign uses this technique.
NetFlow-based deanonymization and puzzle side-channel — 2026 preprint
The October 10, 2026 paper On the Effectiveness of NetFlow-Based Deanonymization of Tor Onion Services studies active traffic-pattern induction and a passive technique using public puzzle-difficulty changes from Tor's DoS defense. It reports a 96.4% accuracy result in a synchronized-DoS evaluation involving 55 targets and a perfect-knowledge candidate-subset condition. That number is not a general-world success rate for arbitrary Tor users/services. The key relevance is that DoS-related signals can, in principle, create information useful to a separate passive observer.
Interpretation: evidence versus inference
| Indicator | Assessment | What it does / does not establish |
|---|---|---|
| Adaptive multi-vector attacks | Strongly supported | Attack signatures changed over time and after defenses; does not identify objective. |
| Guard/HSDir role-associated disruption | Strongly supported in operator reports | Suggests selective targeting or role-based effects; availability attack remains plausible. |
| Synchronized relay flag loss | Reported observation | Suggests coordination; timestamp and consensus mechanics need caution. |
| Traffic imbalance and altered ratios | Reported measurements | Shows traffic manipulation/load; packet-delay correlation intent is not demonstrated. |
| Staged circuit-creation intensity | Reported measurements | Could be adaptive rate control or probing; no discrete payload-size progression documented. |
| Hot descriptor concentration | Strong operator observation, interpretation uncertain | Indicates concentrated descriptor requests; service identity/intent unknown and hash-ring effects may matter. |
| DoS creates correlation opportunities | Research-supported capability | Independent observers may exploit public signals; does not prove DDoS operator is deanonymizing. |
| Current campaign deanonymized victims | Not publicly demonstrated | No verified victim, correlation result, or attribution was found in the cited public reporting. |
Threat assessment
Subjective probabilities—not statistically derived and not a claim of attribution. The categories are framed as primary explanations and are not fully exclusive in reality.
| Hypothesis | Estimate |
|---|---|
| Primarily availability/capacity disruption, without deliberate deanonymization objective | 45% |
| Targeted disruption of specific relays/services, but no deliberate correlation objective | 25% |
| Deliberate traffic analysis/deanonymization objective with DoS as an enabling mechanism | 30% |
Separately, the likelihood that the activity creates exploitable opportunities for deanonymization—whether for the DoS actor or a different observer—is assessed as materially higher than the probability that the DoS operator itself is running a correlation experiment. The cited research supports this technical possibility, but it does not confirm incident linkage.
What evidence would discriminate DoS from correlation/probing?
- Whether the hot-descriptor set remains stable across attack waves, rotates, or correlates with service activity.
- Attack timing relative to HSDir/Guard acquisition, recovery, and flag loss across independent relays.
- Descriptor-fetch distributions and anonymized descriptor-hash groupings over time, where operators can share safely.
- Correlation of public puzzle-difficulty changes with independently observed traffic/load telemetry, following the research's concern about public side-channel signals.
- Whether escalation steps correlate with specific victim/service activity rather than simply resource saturation, mitigation changes, or attacker rate limits.
- Independent replication and raw counter snapshots with timestamps, software versions, relay role, and baseline windows.
Priority: validate the hot-descriptor and synchronized-role-loss observations across independent operators. Those are currently the most discriminating reported features. Do not infer service identity or publish sensitive descriptor identifiers unless doing so is safe and appropriate for affected operators.
Limitations and corrections
- This is a public-source snapshot, not a Tor Project incident investigation or official attribution.
- Operator posts contain measurements from individual relays; counters are not necessarily comparable across Tor versions, patches, or configurations.
- Some thread authors corrected earlier claims. This report uses later corrections where found, including the bandwidth-authority measurement caveat and the lack of baseline data for some exit-stream interpretations.
- Attacks affecting Guard/HSDir roles can be explained by availability objectives. Role-targeting is not, by itself, proof of deanonymization.
- Research demonstrates capabilities under stated conditions; it does not establish that the campaign has adopted them.
- Exit-relay host compromise is treated as a separate incident unless evidence links it to the DDoS actors.
- Probability estimates are judgmental and should be updated as new evidence appears.
Sources
- Traffic disbalance and overload — September 9 operator post (traffic imbalance, backtrace hypothesis).
- Traffic disbalance and overload — September 10 follow-up (ratio changes and operator assessment).
- Tor 0.4.9.13 release notes, September 23 (security fixes including stream isolation).
- Security Release 0.4.9.13 (Tor Project announcement).
- 333 Tor Exit Relays Became SSH Password Traps (separate host-compromise incident).
- Ebury Tor exits: detection, logs and IOCs (independent incident analysis).
- October 12–13 HSDir attack report (relay-specific measurements and flag timing).
- Tor 0.4.9.15 released — new DoS defenses (October 28).
- BEGIN_DIR/HSDir flood, OOM-kill despite MaxMemInQueues (November 2 operator report).
- Circuit-creation flood since November 6 (primary thread, edits/corrections, hot descriptors and relay-source observations).
- Update on attack in progress post-0.4.17 upgrade (attack-type observations and relay responsiveness).
- From “Onion Not Found” to Guard Discovery (PoPETs 2022 research).
- Descriptors of Exposure: Undermining Tor Anonymity Through Exploiting Descriptor Flood (IEEE S&P 2026).
- On the Effectiveness of NetFlow-Based Deanonymization of Tor Onion Services (October 2026 preprint).