I run cheap IoT cameras the way you're supposed to: in their own zone, zero trust, no lateral access. So when I saw them hammering Chinese cloud endpoints — one of them attempting thousands of uploads a day — sinkholing those domains felt like obvious hygiene. I pointed api.eye4.cn and qiniup.com at 127.0.0.1 and moved on. That was the wrong call, and unwinding it taught me more about DNS than the block ever would have.
The Tempting Block#
The cameras phone home constantly. One was making thousands of upload attempts a day to a Chinese object-storage service. On a paranoid network that pattern screams "block it," and the sinkhole was trivial: a local zone answering those names with a loopback address so the queries die on the Pi and never leave the building. Clean, satisfying, and it broke the cameras.
What It Actually Broke#
These cameras don't stream peer-to-peer. Remote viewing works by cloud relay — the camera and the phone app both connect out to the vendor's infrastructure, and the cloud brokers the session. The handshake that sets that up runs through api.eye4.cn. Sinkhole that name and you haven't hardened the camera; you've cut its remote viewing off at the knees. And cheap cameras do not fail gracefully. They don't back off. They retry — immediately, forever — which turns one bad block into a traffic problem.
The False Alarm#
Before I understood that, the DNS logs handed me a red herring. It looked like one camera's queries to the storage domain were escaping the sinkhole — some answers were coming back as real upstream records instead of loopback. I went looking for a wildcard leak. There wasn't one. The sinkhole had only ever covered a couple of specific names plus the apex, never a wildcard, so plenty of subdomains were always going to resolve normally. What made it look broken was stale cache: the resolver was serving a mix of old and new answers. A cache flush made the picture honest again. Lesson one: before you chase a leak, make sure you're not just looking at cache.
The Retry Storm and the TC Bit#
Here's the part I'll remember. One camera sat pinned at over a hundred queries per minute, all day, asking for the same handful of names roughly once a second. The cause was a per-minute query-rate limiter on the resolver. When a client trips the limit, the resolver doesn't drop the query — it answers with an empty response that has the truncation bit set. In a correct DNS client, a truncated answer means "this didn't fit over UDP, ask me again over TCP." A tcpdump on the wire showed exactly that: empty answers, zero records, the TC flag lit. The camera's firmware ignores the truncation bit entirely. It never retries over TCP. It just fires another UDP query — and another, and another — which is precisely the loop the rate limiter was supposed to stop. The limiter and the broken client were feeding each other.
Source-IP Is the Real Boundary#
The takeaway that reshaped my camera rules: you cannot meaningfully filter these devices on destination ports. The relay protocol sprays across a wide, shifting range of ports — too many to enumerate, and they change. Trying to allow "just the right ports" is a losing game that ends with broken video or an unmaintainable list. The boundary that actually holds is the source. The camera's own IP, grouped and scoped, is the security primitive. Let the camera reach its cloud, let the cloud domains resolve, and put the wall around which device is allowed to do that — not which ports it uses.
The Rule I Wrote on the Wall#
So the blocks came back off, permanently, with a note to my future self not to re-add them. The camera cloud domains — the iCSee API, the storage backend, the other vendors' relay and clip-upload endpoints — all resolve normally now, by design. And the high query rate I was so alarmed by? That's just what this firmware does: hardcoded round-robin connectivity checks, retried aggressively. It is normal behavior, not an attack, and the right response is to recognize it rather than sinkhole it. Real isolation here was never about starving the cameras of DNS. It was about making sure that when one of them gets compromised, the only place it can reach is the cloud it was always going to talk to — and nothing else on my network.