6 ms·

On the Quad9 website: >Since Quad9 already performs DNSSEC validation, DNSSEC being enabled in the forwarder will cause a duplication of the DNSSEC process, si

On the Quad9 website: >Since Quad9 already performs DNSSEC validation, DNSSEC being enabled in the forwarder will cause a duplication of the DNSSEC process, significantly reducing performance and potentially causing false BOGUS responses. This sounds dodgy. Surely that means Quad9 can poison my DNS?

Arainach 12d ago
Not if Quad9 is using DNSSEC, no. What's the specific threat you're envisioning? If it involves Quad9 themselves being malicious, what would DNSSEC on the forwarding prevent? This page explains how all of this works in detail: https://quad9.net/news/blog/quad9-enables-dnssec-on-all-service-endpoints/ https://quad9.net/news/blog/quad9-enables-dnssec-on-all-serv...
- zamadatix 12d ago
  DNSSEC validation on your forwarder would prevent a maliciously modified record from Quad9 (or others) from being accepted locally - i.e. "Quad9 can poison my DNS". I've always been of two minds on this. On one hand, that concern is beyond any reasonable level of security/performance/reliability tradeoff for most any user. At the same time, it is a bit of a shame DNS doesn't have a more scalable & performant approach to security which can just always be done without having to consider it a tradeoff, however minor in practice.
  - justsomehnguy 12d ago
    You are rely here on the assumption what your resolver already knows what the zone is DNSSEC signed. If your forwarder or resolver strips that information?..
    - zamadatix 12d ago
      At a high level, one of 3 things happens: 1. The forwarder gets a response claiming the record is supposed to be DNNSEC signed from the parent (recursively traversing from the root). The forwarder checks the signature of this claim. If the signature is valid, the forwarder continues on to validate the signature of the record and checks its validity to know if the info was secure. If the signature is invalid, the forwarder knows any information any information is not able to be validated as secure - Somewhere during the recursive checks through the root, the forwarder gets an unsigned (no DNSSEC) or invalidly signed (e.g. your stripped response) response. The forwarder knows any information is not able to be validated as secure. - A claim for lack of configuration or support of DNSSEC records comes back. The forwarder knows any information is not able to be validated as secure. So you always know whether or not the information was secure, it's just if it was insecure you don't know if it's because it was just never secured or if someone tried to tamper with it. And that should make sense, an insecure message is by definition one which you can't tell if it has been tampered with.
      - tptacek 10d ago
        You have two recursers (or recurser-capable systems, in the case of a forwarder), an upstream that you tentatively trust and a downstream one you trust completely (because end systems use stub resolvers and have to blindly trust a recursive resolver somewhere). Can you be more specific about how, using current DNS software, that downstream resolver can reliably detect whether a zone is DNSSEC signed? The upstream recurser can erase every DNSSEC record in the tree if it wants. What does the downstream recurser do short of jettisoning the upstream and doing all its own recursive lookups?
        zamadatix 10d ago
        I can sense the doubt in the questions but I assure you what I'm discussing is just the defaults behavior of most DNS server software. E.g., I'll walk through below with BIND's default behavior when configured with recursion and a forwarder entry. The downstream forwarder (our BIND server) asks for all of the relevant DNSSEC records for the query to "example.com." from the root on down the FQDN. The upstream responds with a set of record responses. One of 3 things happens: 1) The signature chain checks out all the way to the root, who's key info comes from a file included the DNS software rather than a DNSSEC query. The integrity of the end record is trusted (not the upstream itself) and the process is complete. 2) The signature chain checks out but part way down it (explicitly) ends early. E.g. "example" is not configured for DNSSEC but "com" and "." are so .com ends the signature chain with a signed "example will not be signed" handoff. The process is complete, the record is securely known to be insecurely defined by the authoritative servers for "example" per the signatures from the .com records and the local security anchor (e.g. bind.keys file) 3) The signature chain doesn't validate at some point. This could be because the malicious upstream tried to modify a record or it could be because the malicious upstream tried to fake lack of record response even though it can't fake the parent domain's signature of that (recursively to the root because the root keys are local to the server). In this case, the invalid records from the malicious upstream are jettisoned in favor of trying to get valid responses via recursive resolution. . Tying it all back together: Because the downstream forwarder is only looking for the signature chain to trust the record, it just wants to get those records as fast as possible. BIND determines the most likely way to get a quick response is to ask the forwarder. It then checks the full signature of all of these, almost always bind was right - the forwarder had the records needed cached and didn't try to modify them. If the forwarder was malicious or just in error bind throws these away and tries to recurse the answer directly. Quad9 wants people to disable it because it's extra record requests, most people don't care, and, if you trust Quad9, you can just send your request in a DoH or DoT tunnel to secure the last hop more efficiently while they do the validation of the record itself.
- tptacek 12d ago
  Wait, I must be misunderstanding you, because if you're resolving off Quad9, they can definitely poison your DNSSEC-signed records. Between a stub resolver and a recursor DNSSEC collapses down to a single "yes it was signed" bit in the header. To protect yourself from an upstream resolver using DNSSEC, you need to be doing something akin to a full recursive lookup yourself. This is a flaw in the DNSSEC design and a reason why DoH took off instead.
  - QDwQ1 12d ago
    > To protect yourself from an upstream resolver using DNSSEC, you need to be doing something akin to a full recursive lookup yourself. This is a flaw in the DNSSEC design and a reason why DoH took off instead. What? I don't see how you can call that a flaw in DNSSEC when DoH is no better in this regard; it doesn't even attempt to protect against a malicious recursive resolver. The only way to do that is to validate DNSSEC on the client.
    - tptacek 12d ago
      DNSSEC and DoH provide different security services. But to get the benefit of DNSSEC, you need to resolve recursively. DoH works for stub resolvers. That's all I'm saying.
  - joveian 12d ago
    I don't know the details but knot-resolver asks for DS records of the domain being looked up when forwarding to avoid needing to do a full recursive lookup to validate DNSSEC. As I recall from what I've read this works almost everywhere, including Quad9, but not Google DNS due to a bug that they claim to have identified three years ago but haven't fixed. But as I understand it this is not how recursive resolvers validate so it is extra data that the forwarding server needs to request and cache.
thenewnewguy 12d ago
Technically yes, in practice the odds your local resolver is validating DNSSEC is slim (and if you're intentionally configured it to do so, switch to a provider that isn't Quad9).
usr1106 12d ago
I use Clouldflare DoT and enabled validation in systemd-resolved some time ago. Not because I would be particularly paranoid, but more out of curiosity how it works. I noticed no problems, except for with Atlassian. They use 2 second level domains (at least), one under .com signed and one under .net (unsigned) (IIRC). Most things worked like normal, but some Jira extension stopped working. Turned out that systemd did reject their signed subdomains. Could not figure out whether the rejection was justified or not. When I asked Claude 7 times about it, I also got 7 contradicting answers... Reported to Atlassian support that their signing is incorrect (some delegation missing). To my surprise they replied: Are you using systemd? And gave a bug number that systemd handles validation wrong in their case. Haven't had time to study the bug and really understand the whole issue myself. Just left it there with the takeaway that local validation is currently not for non-experts. (Sorry not at my computer. Details rather vague from memory.)

Arainach 12d ago

Not if Quad9 is using DNSSEC, no. What's the specific threat you're envisioning? If it involves Quad9 themselves being malicious, what would DNSSEC on the forwarding prevent? This page explains how all of this works in detail: https://quad9.net/news/blog/quad9-enables-dnssec-on-all-service-endpoints/ https://quad9.net/news/blog/quad9-enables-dnssec-on-all-serv...

zamadatix 12d ago
DNSSEC validation on your forwarder would prevent a maliciously modified record from Quad9 (or others) from being accepted locally - i.e. "Quad9 can poison my DNS". I've always been of two minds on this. On one hand, that concern is beyond any reasonable level of security/performance/reliability tradeoff for most any user. At the same time, it is a bit of a shame DNS doesn't have a more scalable & performant approach to security which can just always be done without having to consider it a tradeoff, however minor in practice.
- justsomehnguy 12d ago
  You are rely here on the assumption what your resolver already knows what the zone is DNSSEC signed. If your forwarder or resolver strips that information?..
  - zamadatix 12d ago
    At a high level, one of 3 things happens: 1. The forwarder gets a response claiming the record is supposed to be DNNSEC signed from the parent (recursively traversing from the root). The forwarder checks the signature of this claim. If the signature is valid, the forwarder continues on to validate the signature of the record and checks its validity to know if the info was secure. If the signature is invalid, the forwarder knows any information any information is not able to be validated as secure - Somewhere during the recursive checks through the root, the forwarder gets an unsigned (no DNSSEC) or invalidly signed (e.g. your stripped response) response. The forwarder knows any information is not able to be validated as secure. - A claim for lack of configuration or support of DNSSEC records comes back. The forwarder knows any information is not able to be validated as secure. So you always know whether or not the information was secure, it's just if it was insecure you don't know if it's because it was just never secured or if someone tried to tamper with it. And that should make sense, an insecure message is by definition one which you can't tell if it has been tampered with.
    - tptacek 10d ago
      You have two recursers (or recurser-capable systems, in the case of a forwarder), an upstream that you tentatively trust and a downstream one you trust completely (because end systems use stub resolvers and have to blindly trust a recursive resolver somewhere). Can you be more specific about how, using current DNS software, that downstream resolver can reliably detect whether a zone is DNSSEC signed? The upstream recurser can erase every DNSSEC record in the tree if it wants. What does the downstream recurser do short of jettisoning the upstream and doing all its own recursive lookups?
      - zamadatix 10d ago
        I can sense the doubt in the questions but I assure you what I'm discussing is just the defaults behavior of most DNS server software. E.g., I'll walk through below with BIND's default behavior when configured with recursion and a forwarder entry. The downstream forwarder (our BIND server) asks for all of the relevant DNSSEC records for the query to "example.com." from the root on down the FQDN. The upstream responds with a set of record responses. One of 3 things happens: 1) The signature chain checks out all the way to the root, who's key info comes from a file included the DNS software rather than a DNSSEC query. The integrity of the end record is trusted (not the upstream itself) and the process is complete. 2) The signature chain checks out but part way down it (explicitly) ends early. E.g. "example" is not configured for DNSSEC but "com" and "." are so .com ends the signature chain with a signed "example will not be signed" handoff. The process is complete, the record is securely known to be insecurely defined by the authoritative servers for "example" per the signatures from the .com records and the local security anchor (e.g. bind.keys file) 3) The signature chain doesn't validate at some point. This could be because the malicious upstream tried to modify a record or it could be because the malicious upstream tried to fake lack of record response even though it can't fake the parent domain's signature of that (recursively to the root because the root keys are local to the server). In this case, the invalid records from the malicious upstream are jettisoned in favor of trying to get valid responses via recursive resolution. . Tying it all back together: Because the downstream forwarder is only looking for the signature chain to trust the record, it just wants to get those records as fast as possible. BIND determines the most likely way to get a quick response is to ask the forwarder. It then checks the full signature of all of these, almost always bind was right - the forwarder had the records needed cached and didn't try to modify them. If the forwarder was malicious or just in error bind throws these away and tries to recurse the answer directly. Quad9 wants people to disable it because it's extra record requests, most people don't care, and, if you trust Quad9, you can just send your request in a DoH or DoT tunnel to secure the last hop more efficiently while they do the validation of the record itself.
tptacek 12d ago
Wait, I must be misunderstanding you, because if you're resolving off Quad9, they can definitely poison your DNSSEC-signed records. Between a stub resolver and a recursor DNSSEC collapses down to a single "yes it was signed" bit in the header. To protect yourself from an upstream resolver using DNSSEC, you need to be doing something akin to a full recursive lookup yourself. This is a flaw in the DNSSEC design and a reason why DoH took off instead.
- QDwQ1 12d ago
  > To protect yourself from an upstream resolver using DNSSEC, you need to be doing something akin to a full recursive lookup yourself. This is a flaw in the DNSSEC design and a reason why DoH took off instead. What? I don't see how you can call that a flaw in DNSSEC when DoH is no better in this regard; it doesn't even attempt to protect against a malicious recursive resolver. The only way to do that is to validate DNSSEC on the client.
  - tptacek 12d ago
    DNSSEC and DoH provide different security services. But to get the benefit of DNSSEC, you need to resolve recursively. DoH works for stub resolvers. That's all I'm saying.
- joveian 12d ago
  I don't know the details but knot-resolver asks for DS records of the domain being looked up when forwarding to avoid needing to do a full recursive lookup to validate DNSSEC. As I recall from what I've read this works almost everywhere, including Quad9, but not Google DNS due to a bug that they claim to have identified three years ago but haven't fixed. But as I understand it this is not how recursive resolvers validate so it is extra data that the forwarding server needs to request and cache.

thenewnewguy 12d ago

Technically yes, in practice the odds your local resolver is validating DNSSEC is slim (and if you're intentionally configured it to do so, switch to a provider that isn't Quad9).

usr1106 12d ago

I use Clouldflare DoT and enabled validation in systemd-resolved some time ago. Not because I would be particularly paranoid, but more out of curiosity how it works. I noticed no problems, except for with Atlassian. They use 2 second level domains (at least), one under .com signed and one under .net (unsigned) (IIRC). Most things worked like normal, but some Jira extension stopped working. Turned out that systemd did reject their signed subdomains. Could not figure out whether the rejection was justified or not. When I asked Claude 7 times about it, I also got 7 contradicting answers... Reported to Atlassian support that their signing is incorrect (some delegation missing). To my surprise they replied: Are you using systemd? And gave a bug number that systemd handles validation wrong in their case. Haven't had time to study the bug and really understand the whole issue myself. Just left it there with the takeaway that local validation is currently not for non-experts. (Sorry not at my computer. Details rather vague from memory.)