If LE hacked the SR server they could do traffic analysis and link anyone who uses their entry guards to individual accounts without cookies. I am not sure that the application layer attack is really that much worse than network layer attacks. Certainly application layer attacks are more in our control to defend against, and network layer attacks are statistical probabilities over time. But both are serious threats. The feds got some subset of people who visited FH by hacking them from FH, but the feds could also get some subset of people who visited FH by owning their entry guards and doing end point traffic correlation. For all we know they did both in this case, it is just easier to identify application layer attacks than it is to identify traffic analysis. And isolation etc could have protected people from the application layer attack, but Tor itself is totally incapable of protecting people from correlation attacks if they have a bad entry node and go to a compromised server, at best it can decrease the probability that the victim has a bad entry guard by getting more good users to run relays. But what if the NSA does a passive attack and feeds the intelligence to DEA? Then it no longer matters if your entry guards are good or not if you are in USA or your entry guards are or your traffic passes through USA on the way to your entry guards. In the past I put more faith in Tor than I currently do, and was more worried about application layer attacks. And I did think the feds would do application layer attacks prior to traffic analysis attacks, and was apparently correct about it. Now I am worried about traffic analysis and application layer attacks, and I bet the feds start using both. Application Attacks: Easier to add defenses that mitigate, theoretically possible but unrealistic to fully protect from Traffic Analysis: Harder to add defenses that mitigate, theoretically possible but unrealistic to fully protect from Application Attacks: More likely to be noticed Traffic Analysis: Much less likely to be noticed Application Attacks: Capable of taking full control of remote system and stealing private keys, plaintexts, etc Traffic Analysis: Only capable of obtaining suspect IP address to a high degree of certainty Application Attacks: Constantly evolving threat with no end in sight, new zero days all the time thousands and thousands waiting to be discovered, attacks are fully protected from shortly after they are discovered Traffic Analysis: Largely understood, slowly evolving with few new attacks, old attacks are rarely able to be fully protected from Application Attacks: Security advances are making application attacks more and more difficult Traffic Analysis: Passive surveillance is making traffic analysis harder and harder to protect from Application Attacks: Are more likely to deanonymize all *vulnerable* users immediately Traffic Analysis: Is more likely to slowly deanonymize *ALL* users over time Application Attacks: Are more likely to target a subset of users rather than all users, but likely to compromise all targeted users Traffic Analysis: Is more likely to target all users but only compromise a subset of targeted users Application Attacks: Are trivial and cheap to do against users who do not stay on their toes and keep fully patched Traffic Analysis: Is not usually easier to do against users who are not fully patched, but it can be (ie: the introduction of guards) Application Attacks: Are expensive to do against users who stay fully patched and very expensive to do against users who stay fully patched and use layers of isolation and other defense mechanisms, cost increases substantially as subset of users to target increases. Traffic Analysis: Can be made more expensive to do in some cases but there is a hard and low ceiling tied to the anonymity technology being used, is usually roughly as effective against all users regardless of their configuration (with some variance but not nearly as much as compared to application attacks), cost correlates directly with time, the more the attacker spends the less time they need to wait to deanonymize their targets, the less they spend the longer they need to wait Application Attacks: Quickly identify all vulnerable users but become less effective over time as users patch and awareness spreads Traffic Analysis: Identifies targets with various speed depending on amount spent on it, the more time that passes the more targets are identified Application Attacks: Have a one time cost to obtain but become less valuable as time passes Traffic Analysis: Has continuous cost to maintain but becomes more valuable as time passes