Signals, Not Forts
The review finding was correct. Our service accepted location and certificate headers that the browser could set for itself. Anyone could write “I am in this city” or “I presented this certificate” into a request, and the service would carry it along as a fact about the caller.
The fix was obvious: strip them. A header the client can forge is a header nobody should trust, so the door removes it before anything downstream sees it. The change was small, and the finding was closed.
Then the operator read the change and reversed it. The paraphrase is close to the original: having those headers is a valid abuse signal, so send everything that is not outright bad, and in every case record that it happened.
What the strip destroyed
A forged header is not noise. It is the one piece of evidence that somebody is trying.
Consider what the strip leaves behind. The request arrives looking clean. The service has no record that the browser claimed to be somewhere it was not, or that a certificate was asserted by a client with no way to hold one. The attempt happened, and the system’s only response was to make it invisible. An abuse analyst, a fraud team or a later audit now has nothing to mine, because the thing worth mining was deleted at the door.
That is the difference in kind between a fort and a gyroscope. A fort drops what it distrusts and carries on as if nothing had come. A gyroscope notices the disturbance, keeps its orientation, and keeps the reading. (An earlier post on the gyroscope and the wall made the general case. This is one concrete place it applied.)
The mechanism: forward, grade, record
The rule that came out of the reversal has three parts.
Forward everything except the outright bad. Outright bad means a credential travelling where it must not, input that is malformed or oversized, or something known to be malicious. Those are refused. Everything else goes through.
Grade every value by who stands behind it. There are four grades, in plain terms:
- edge-witnessed: written by infrastructure we operate, in front of the service;
- relayed by a checked forwarder: supplied by a service that the receiver has itself checked, and honoured only from that sender;
- client-asserted: the browser said so, and nobody stood behind it;
- self-declared: a downstream service describing itself.
The grade, not the presence of the value, decides what the value may do. Policy may use only what the grade entitles it to. A client-asserted location never feeds a policy decision. In the code that is a property of the types: the asserted values live in a distinct structure from the graded ones, so a client claim cannot be ranked against an edge fact by accident. The analysis side, by contrast, reads everything.
Record every drop by name and reason. The blocked list carries the header’s name and a reason code (credential-shaped, oversized, malformed, past the cap), and never its value. A claim that exceeds a limit is blocked and recorded, not silently truncated into something that looks accepted.
Where the signal comes from
The forwarded client assertions travel in one bounded envelope, separate from the edge-witnessed facts. When the edge has witnessed a fact and the client also asserted it, the two are compared. If they differ, a typed spoof signal is raised, naming which facts disagreed. If the edge witnessed nothing, there is nothing to compare, and no signal is raised. The service does not guess.
The interface shows the same separation. Browser-claimed values render apart, in their own box, marked as untrusted and client-asserted, with any spoof signal beside them and a count of what was blocked. They are never merged into the row that shows what the edge saw.
This is the same stance the sandbox in our scan pipeline already took: it declaws what it receives, but does not judge it, and anything downstream can revoke a verdict later. The doctrine only named it.
What changed
The reversal is recorded as a standing decision, not as a change to one file. That matters, because the strip was an instinct a reviewer will have again.
- The forgeable headers are forwarded, graded client-asserted, and never used for policy.
- Every refusal produces a record with a name and a reason.
- A mismatch between what the edge saw and what the client said is a typed signal.
The change is built, recorded in the design, and shipped in the following release.
For the longer account of what happens to a record once it exists, see a record for every attempt. This post is the step before it: not destroying the attempt in the first place.
The line that stays
A fort drops; a gyroscope grades.