MadeYouReset works by making the server cancel its own HTTP/2 streams, rather than by having the client cancel them as HTTP/2 Rapid Reset did. It is an HTTP/2 denial-of-service attack, tracked as CVE-2025-8671 (CVSS 7.5) and disclosed on 13 August 2025 by academic and industry researchers, and it is one of the protocol-layer vectors a DDoS test checks at every hop that speaks HTTP/2. The reset still happens, the stream still stops counting, and the request behind it keeps running. Only the sender of the reset has changed, and that is enough to walk past the limits built in 2023.
How MadeYouReset works
The Rapid Reset patches mostly counted one thing: RST_STREAM frames arriving from the client. A connection whose client cancelled an implausible share of its streams was closed. That shut the obvious door.
HTTP/2 has another way for a stream to end early. When a frame on an open stream breaks the protocol's rules in a way that affects only that stream, the specification has the server answer with a stream error, which it delivers as its own RST_STREAM. The connection stays up. The stream is closed.
MadeYouReset sends malformed frames chosen to trigger exactly that. Each open request gets one, and the server resets the stream on the client's behalf.
The stream slot is freed the moment the server's own reset goes out, so the client's concurrency budget refills as fast as it can provoke errors. The request behind that slot is a different matter. Once handed to the application it is not recalled, and the server keeps working on requests the client no longer counts. The client gets the effect of Rapid Reset without ever sending a reset.
Why it matters under attack
The mitigation is looking the wrong way. A reset counter that only watches inbound RST_STREAM frames reads zero for the whole attack. Dashboards built after 2023 report a quiet connection while backend work piles up.
The work is real and invisible. Each reset stream leaves a request running that nobody is waiting for. Concurrency limits see open streams, not running handlers, so the gap between the two grows until workers, memory or upstream connections run out for real users.
The 2023 patch level proves little. Many stacks fixed for CVE-2023-44487 needed a second fix for this one, so an inventory that marks a hop "Rapid Reset: patched" says nothing about CVE-2025-8671 on the same box.
Across the family, each variant hides the request at a different moment: the HTTP/2 CONTINUATION flood before its headers ever finish, MadeYouReset after the protocol has already closed it.
Hardening, and what a DDoS test measures
The fix comes in two parts. Run HTTP/2 stacks patched for CVE-2025-8671 at every terminating hop. And make the reset limit count resets in both directions: a connection on which the server keeps resetting streams because of client errors is as suspicious as one on which the client keeps cancelling.
The test question differs from the Rapid Reset one in a single word: not "does a hop close the connection after too many cancels" but "does it close it after too many errors it answered itself". The reading is the count of server-issued resets a connection is allowed before it ends, set beside the client-reset limit on the same hop. Two numbers that disagree by orders of magnitude are the finding.
Backend timing completes it. Compare stream-close times at the HTTP/2 hop with request-completion times further in; a growing band of work finishing after its stream died is the cost, and the goodput real sessions get during that window is what it costs users. How these request-layer vectors fit beside HTTP floods and slow attacks in one plan is covered in application-layer DDoS testing.