Having spent the last 36 hours crunching ICANN’s lists of almost 10 million new gTLD name collisions, the DI PRO collisions database is back online, and we can start reporting some interesting facts.
First, while we reported yesterday that 1,318 new gTLD applicants will be asked to block a total of 9.8 million unique domain names, the number of distinct second-level strings involved is somewhat smaller.
It’s 6,806,050, according to our calculations, still a bewilderingly high number.
The most commonly blocked string, as expected, is “www”. It’s on the block-lists for 1,195 gTLDs, over 90% of the total.
Second is “2010”. I currently have no explanation for this, but I’m wondering if it’s an artifact of the years of Day In The Life data upon which ICANN based its lists.
Protocol-related strings such as “wpad” and “isatap” also rank highly, as do strings matching popular TLDs such as “com”, “org”, “uk” and “de”. Single-character strings are also very popular.
The brand with the most blocks (free trademark protection?) is unsurprisingly Google.
The string “google” appears as an exact match on 930 gTLDs’ lists. It appears as a substring of 1,235 additional blocked strings, such as “google-toolbar” and “googlemaps”.
Facebook, Yahoo, Gmail, YouTube and Hotmail also feature in the top 100 blocked brands.
DI PRO subscribers can search for strings that interest them, discovering how many and which gTLDs they’re blocked in, using the database.
Here’s a table of the top 50 blocked strings.
With all the talk recently about DNS Namespace Collisions, the heretofore relatively obscure Day In The Life (“DITL”) datasets maintained by the DNS-OARC have been getting a lot of attention.
While these datasets are well known to researchers, I’d like to take the opportunity to provide some background and talk a little about how these datasets are being used to research the DNS Namespace Collision issue.
The Domain Name System Operations Analysis and Research Center (“DNS-OARC”) began working with the root server operators to collect data in 2006. The effort was coined “Day In The Life of the Internet (DITL).”
Root server participation in the DITL collection is voluntary and the number of contributing operators has steadily increased; in 2010, all of the 13 root server letters participated. DITL data collection occurs on an annual basis and covers approximately 50 contiguous hours.
DNS-OARC’s DITL datasets are attractive for researching the DNS Namespace Collision issue because:
- DITL contains data from multiple root operators;
- The robust annual sampling methodology (with samples dating back to 2006) allows trending; and
- It’s available to all DNS-OARC Members.
More information on the DITL collection is available on DNS-OARC’s site at https://www.dns-oarc.net/oarc/data/ditl.
Terabytes and terabytes of data
The data consists of the raw network “packets” destined for each root server. Contained within the network packets are the DNS queries. The raw data consists of many terabytes of compressed network capture files and processing the raw data is very time-consuming and resource-intensive.
While several researchers have looked at DITL datasets over the years, the current collisions-oriented research started with Roy Hooper of Demand Media. Roy created a process to iterate through this data and convert it into intermediate forms that are much more usable for researching the proposed new TLDs.
We started with his process and continued working with it; our code is available on GitHub for others to review.
Finding needles in DITL haystacks
The first problem faced by researchers interested in new TLDs is isolating the relatively few queries of interest among many terabytes of traffic that are not of interest.
Each root operator contributes several hundred – or several thousand – files full of captured packets in time-sequential order. These packets contain every DNS query reaching the root that requests information about DNS names falling within delegated and undelegated TLDs.
The first step is to search these packets for DNS queries involving the TLDs of interest. The result is one file per TLD containing all queries from all roots involving that TLD. If the input packet is considered a “horizontal” slice of root DNS traffic, then this intermediary work product is a “vertical” slice per TLD.
These intermediary files are much more manageable, ranging from just a few records to 3 GB. To support additional investigation and debugging, the intermediary files that JAS produces are fully “traceable” such that a record in the intermediary file can be traced back to the source raw network packet.
The DITL data contain quite a bit of noise, primarily DNS traffic that was not actually destined for the root. Our process filters the data by destination IP address so that the only remaining data is that which was originally destined for the root name servers.
JAS has made these intermediary per-TLD files available to DNS-OARC members for further analysis.
The intermediary files are comparatively small and easy to parse, opening the door to more elaborate research. For example, JAS has written various “second passes” that classify queries, separate queries that use valid syntax at the second level from those that don’t, detect “randomness,” fit regular expressions to the queries, and more.
We have also checked to confirm that second level queries that look like Punycode IDNs (start with ‘xn--‘) are valid Punycode. It is interesting to note the tremendous volume of erroneous, technically invalid, and/or nonsensical DNS queries that make it to the root.
Also of interest is that the datasets are dominated by query strings that appear random and/or machine-generated.
Google’s Chrome browser generates three random 10-character queries upon startup in an effort to detect network properties. Those “Chrome 10” queries together with a relatively small number of other common patterns comprise a significant proportion of the entire dataset.
Research is being done in order to better understand the source of these machine-generated queries.
More technical details and information on running the process is available on the DNS-OARC web site.
This is a guest post written by Kevin White, VP Technology, JAS Global Advisors LLC. JAS is currently authoring a “Name Collision Occurrence Management Framework” for the new gTLD program under contract with ICANN.
Verisign has rubbished the Commonwealth Bank of Australia’s claim that its dot-brand gTLD, .cba, is safe.
In a lengthy letter to ICANN today, Verisign senior vice president Pat Kane said that, contrary to CBA’s claims, the bank is only responsible for about 6% of the traffic .cba sees at the root.
It’s the latest volley in the ongoing fight about the security risks of name collisions — the scenario where an applied-for gTLD string is already in broad use on internal networks.
CBA’s application for .cba has been categorized as “uncalculated risk” by ICANN, meaning it faces more reviews and three to six months of delay while its risk profile is assessed.
But in a letter to ICANN last month, CBA said “the cause of the name collision is primarily from CBA internal systems” and “it is within the CBA realm of control to detect and remediate said systems”.
The bank was basically claiming that its own computers use DNS requests for .cba already, and that leakage of those requests onto the internet was responsible for its relatively high risk profile.
At the time we doubted that CBA had access to the data needed to draw this conclusion and Verisign said today that a new study of its own “shows without a doubt that CBA’s initial conclusions are incorrect”.
Since the publication of Interisle Consulting’s independent review into root server error traffic — which led to all applied-for strings being split into risk categories — Verisign has evidently been carrying out its own study.
While Interisle used data collected from almost all of the DNS root servers, Verisign’s seven-week study only looked at data gathered from the A-root and J-root, which it manages.
According to Verisign, .cba gets roughly 10,000 root server queries per day — 504,000 in total over the study window — and hardly any of them come from the bank itself.
Most appear to be from residential apartment complexes in Chiba, Japan, where network admins seem to have borrowed the local airport code — also CBA — to address local devices.
About 80% of the requests seen come from devices using DNS Service Discovery services such as Bonjour, Verisign said.
Bonjour is an Apple-created technology that allows computers to use DNS to automatically discover other LAN-connected devices such as printers and cameras, making home networking a bit simpler.
Another source of the .cba traffic is McAfee’s antivirus software, made by Intel, which Verisign said uses DNS to check whether code is virus-free before executing it.
While error traffic for .cba was seen from 170 countries, Verisign said that Japan — notable for not being Australia — was the biggest source, with almost 400,000 queries (79% of the total). It said:
Our measurement study reveals evidence of a substantial Internet-connected infrastructure in Japan that lies beneath the surface of the public-facing internet, which appears to rely on the non-resolution of the string .CBA.
This infrastructure appear hierarchical and seems to include municipal and private administrative and service networks associated with electronic resource management for office and residential building facilities, as well as consumer devices.
One apartment block in Chiba is is responsible for almost 5% of the daily .cba queries — about 500 per day on average — according to Verisign’s letter, though there were 63 notable sources in total.
ICANN’s proposal for reducing the risk of these name collisions causing problems would require CBA, as the registry, to hunt down and warn organizations of .cba’s impending delegation.
Verisign reiterates the point made by RIPE NCC last month: this would be quite difficult to carry out.
But it does seem that Verisign has done a pretty good job tracking down the organizations that would be affected by .cba being delegated.
The question that Verisign’s letter and presentation does not address is: what would happen to these networks if .cba was delegated?
If .cba is delegated, what will McAfee’s antivirus software do? Will it crash the user’s computer? Will it allow unsafe code to run? Will it cause false positives, blocking users from legitimate content?
Or will it simply fail gracefully, causing no security problems whatsoever?
Likewise, what happens when Bonjour expects .cba to not exist and it suddenly does? Do Apple computers start leaking data about the devices on their local network to unintended third parties?
Or does it, again, cause no security problems whatsoever?
Without satisfactory answers to those questions, maybe name collisions could be introduced by ICANN with little to no effect, meaning the “risk” isn’t really a risk at all.
Answering those questions will of course take time, which means delay, which is not something most applicants want to hear right now.
Verisign’s study targeted CBA because CBA singled itself out by claiming to be responsible for the .cba error traffic, not because CBA is a client of rival registry Afilias.
The bank can probably thank Verisign for its study, which may turn out to be quite handy.
Still, it would be interesting to see Verisign conduct a similar study on, say, .windows (Microsoft), .cloud (Symantec) or .bank (Financial Services Roundtable), which are among the 35 gTLDs with “uncalculated” risk profiles that Verisign promised to provide back-end registry services for before it decided that new gTLDs were dangerous.
New gTLDs are set to be added to the widely used Public Suffix List within a month of signing an ICANN registry agreement, according to PSL volunteer Jothan Frakes.
This is pretty good news for new gTLD registries.
The PSL, maintained by volunteers under the Mozilla banner, is used in browsers including Firefox and Chrome, and will be a vital part of making sure new gTLDs “work” out of the box.
If a TLD doesn’t have an entry on the PSL, browsers tend to handle them badly.
For example, after .sx launched last year, Google’s Chrome browser returned search results instead of the intended web site when .sx domain names were typed into the address/search bar.
It also provides a critical security function, telling browsers at which level they should allow domains to set cookies.
According to Frakes, who has been working behind the scenes with other PSL volunteers and ICANN staff to get this process working, new gTLDs will usually hit the PSL within 30 days of an ICANN contract.
Due to the mandatory pre-delegation testing period, new gTLDs should be on the PSL before or at roughly the same time as they are delegated, with plenty of time to spare before they launch.
The process of being added to the PSL should be fairly quick for TLDs that intend to run flat second-level spaces, according to Frakes, but may be more complex if they plan to do something less standard, such as selling third-level domains, for example.
Browser makers may take some time to update their own lists with the PSL updates. Google, with its own huge portfolio of applications, will presumably be incentivized to stay on the ball.
The first tranche of responses to Interisle Consulting’s study into the security risks of new gTLDs, and ICANN’s proposal to delay a few hundred strings pending more study, is in.
Comments filed with ICANN before the public comment deadline yesterday fall basically into two camps:
- Non-applicants (mostly) urging ICANN to proceed with extreme caution. Many are asking for more time to study their own networks so they can get a better handle on their own risk profiles.
- Applicants shooting holes in Interisle’s study and ICANN’s remeditation plan. They want ICANN to reclassify everything except .home and .corp as low risk, removing delays to delegation and go-live.
They were responding to ICANN’s decision to delay 521 “uncalculated risk” new gTLD applications by three to six months while further research into the risk of name collisions — where a new gTLD could conflict with a TLD already used by internet users in a non-standard way — is carried out.
Proceed with caution
Many commenters stated that more time is needed to analyse the risks posed by name collisions, noting that Interisle studied primarily the volume of queries for non-existent domains, rather than looking deeply into the consequences of delegating colliding gTLDs.
That was a point raised by applicants too, but while applicants conclude that this lack of data should lead ICANN to lift the current delays, others believe that it means more delays are needed.
Two ICANN constituencies seem to generally agree with the findings of the Interisle report.
The Internet Service Providers and Connectivity Providers constituency asked for the public comment period be put on hold until further research is carried out, or for at least 60 days. It noted:
corporations, ISPs and connectivity providers may bear the brunt of the security and customer-experience issues resulting from adverse (as yet un-analyzed) impacts from name collision
these issues, due to their security and customer-experience aspects, fall outside the remit of people who normally participate in the ICANN process, requiring extensive wide-ranging briefings even in corporations that do participate actively in the ICANN process
The At-Large Advisory Committee concurred that the Interisle study does not currently provide enough information to fully gauge the risk of name collisions causing harm.
ALAC said it was “in general concurrence with the proposed risk mitigation actions for the three defined risk categories” anyway, adding:
ICANN must assure that such residual risk is not transferred to third parties such as current registry operators, new gTLD applicants, registrants, consumers and individual end users. In particular, the direct and indirect costs associated with proposed mitigation actions should not have to be borne by registrants, consumers and individual end users. The Board must err on the side of caution
Several individual stakeholders agreed with the ISPCP that they need more time to look at their own networks. The Association of Nation Advertisers said:
Our member companies are working diligently to determine if DNS Clash issues are present within their respective networks. However the ANA had to communicate these issues to hundreds of companies, after which these companies must generate new data to determine the potential service failures on their respective networks.
The ANA wants the public comment period extended until November 22 to give its members more time to gather data.
While the ANA can always be relied upon to ask for new gTLDs to be delayed, its request was echoed by others.
General Electric called for three types of additional research:
- Additional studies of traffic beyond the initial DITL sample.
- Information and analysis of “use cases” — particular types of queries and traffic — and the consequences of the failure of particular use cases to resolve as intended (particular use cases could have severe consequences even if they might occur infrequently — like hurricanes), and
- Studies of the time and costs of mitigation.
GE said more time is needed for companies such as itself to conduct impact analyses on their own internal networks and asked ICANN to not delegate any gTLD until the risk is “fully understood”.
The Association of Competitive Technology (which has Verisign as a member) said:
ICANN should slow or temporarily suspend the process of delegating TLDs at risk of causing problems due to their frequency of appearance in queries to the root. While we appreciate the designation of .home and .corp as high risk, there are many other TLDs which will also have a significant destructive effect.
Numerically, there were far more comments criticizing ICANN’s mitigation proposal. All were filed by new gTLD applicants, whose interests are aligned, however.
Most of these comments, which are far more focused on the details and the data, target perceived deficiencies in Interisle’s report and ICANN’s response to it.
Several very good arguments are made.
The Svalbard problem
First, there is criticism of the cut-off point between “low risk” and “uncalculated risk” strings, which some applicants say is “arbitrary”.
That’s mostly true.
ICANN basically took the list of applied-for strings, ordered by the frequency Interisle found they generate NXDOMAIN responses at the root, and drew a line across it at the 49,842 queries mark.
That’s because 49,842 queries is what .sj, the least-frequently-queried real TLD, received over the same period
If your string, despite not yet existing as a gTLD, already gets more traffic than .sj, it’s classed as “uncalculated risk” and faces more delays, according to ICANN’s plan.
As Directi said in its comments:
The result of this arbitrary selection is that .bio (Rank 281) with 50,000 queries (rounded to the nearest thousand) is part of the “uncategorized risk” list, and is delayed by 3 to 6 months, whereas .engineering (Rank 282) with 49,000 queries (rounded to the nearest thousand) is part of the “low risk” list, and can proceed without any significant delays.
What neither ICANN nor Interisle explained is why this is an appropriate place to draw a line in the sand.
.sj is the ccTLD for Svalbard, a Norwegian territory in the Arctic Circle with fewer than 3,000 inhabitants. The TLD is administered by .no registry Norid, but it’s not possible to register domains there.
Does having more traffic than .sj mean a gTLD is automatically more risky? Does having less mean a gTLD is safe? The ICANN proposal assumes “yes” to both questions, but it doesn’t explain why.
Many applicants say that having more traffic than existing gTLDs does not automatically mean your gTLD poses a risk.
They pointed to Verisign data from 2006, which shows that gTLDs such as .xxx and .asia were already receiving large amounts of traffic prior to their delegation. When they were delegated, the sky did not fall. Indeed, there were no reports of significant security and stability problems.
The New gTLD Applicants Group said:
In fact, the least “dangerous” current gTLD on the chart, .sx, had 331 queries per million in 2006. This is a higher density of NXDOMAIN queries than all but five proposed new TLDs. 4 Again, .sx was launched successfully in 2012 with none of the problems predicted in these reports.
These successful delegations alone demonstrate that there is no need to delay any more than the two most risky strings.
There is no factual basis in the study recommending halting delegation process of 20% of applied-for strings. As the paper itself says, “The Study did not find enough information to properly classify these strings given the short timeline.” Without evidence of actual harm, the TLDs should proceed to delegation. Such was the case with other TLDs such as .XXX and .ASIA, which were delegated without delay and with no problems post-delegation.
Applicants also believe that the release in June 2012 of the list of all 1,930 applied-for strings may have skewed the data set that Interisle used in its study.
Uniregistry, for example, said:
The sole fact that queries are being received at the root level does not itself present a security risk, especially after the release to the public of the list of applied-for strings.
The argument seems to be that a lot of the NXDOMAIN traffic seen in 2013 is due to people and software querying applied-for TLDs to see if they’re live yet.
It’s quite a speculative argument, but it’s somewhat supported by the fact that many applied-for strings received more queries in 2013 than they did in the equivalent 2012 sampling.
Some applicants pointed out that there may not be a correlation between the volume of traffic a string receives and the number of second-level domains being queried.
A string might get a bazillion queries for a single second-level domain name. If that domain name is reserved by the registry, the risk of a name collision might be completely eliminated.
The Interisle report did show that number of SLDs and the volume of traffic do not correlate.
For example, .hsbc is ranked 14th in terms of traffic volume but saw requests for just 2,000 domains, whereas .inc, which ranked 15th, saw requests for 73,000 domains.
Unfortunately, the Interisle report only published the SLD numbers for the top 35 strings by query volume, leaving most applicants none the wiser about the possible impact of their own strings.
And ICANN did not factor the number of SLDs into its decision about where to draw the line between “low” and “uncalculated” risk.
Some applicants questioned whether the Interisle data itself was reliable, but I find these arguments poorly supported and largely speculative.
They propose that someone (meaning presumably Verisign, which stands to lose market share when new gTLDs go live, and which kicked off the name collisions debate in the first place) could have gamed the study by generating spurious requests for applied-for gTLDs during the period Interisle’s data was being captured.
Some applicants put forth this view, while other limited their comments to a request that future studies rely only on data collected before now, to avoid tampering at the point of collection in future.
Query counts are very easily gamed by any Internet connected system, allowing for malicious actors to create the appearance of risk for any string that they may object to in the future. It would be very easy to create the impression of a widespread string collision problem with a home Internet connection and the abuse of the thousands of available open resolvers.
While this kind of mischief is a hypothetical possibility, nobody has supplied any evidence that Interisle’s data was manipulated by anyone.
Some people have privately pointed DI to the fact that Verisign made a substantial donation to the DNS-OARC — the group that collected the data that Interisle used in its study — in July.
The implication is that Verisign was somehow able to manipulate the data after it was captured by DNS-OARC.
I don’t buy this either. We’re talking about a highly complex 8TB data set that took Interisle’s computers a week to process on each pass. The data, under the OARC’s deal with the root server operators, is not allowed to leave its premises. It would not be easily manipulated.
Additionally, DNS-OARC is managed by Internet Systems Consortium — which runs the F-root and is Uniregistry’s back-end registry provider — from its own premises in California.
In short, in the absence of any evidence supporting this conspiracy theory, I find the idea that the Interisle data was hacked after it was collected highly improbable.
Several ICANN constituencies that can usually be relied upon to comment on everything (registrars, intellectual property, business and non-commercial) have not yet commented.
Will ICANN extend the deadline? I suppose it depends on how cautious it wants to be, whether it believes the companies requesting the extension really are conducting their own internal collision studies, and how useful it thinks those studies will be.