HttpArena
Knowledge BaseFrameworks

Resilience: The 41 Attacks

The suite has 41 attacks across four protocol surfaces. The most valuable for this benchmark are the server-level ones that fire regardless of what the handler does — most entries are minimal handlers that never parse the body, so the parser-differential attacks (the POST /parse table below) mostly read survived on them, which is itself a recordable result. The Exhausts column says what each attack targets; how that becomes a verdict is read from the container afterwards.

HTTP/1.1 server-level

Fire against any entry:

AttackExhaustsWhat it does
slow-bodymemoryPOST /echo, 100 MiB declared Content-Length, body dripped ~256 B/s
slow-headersconnectionsSlowloris — an unfinished, ever-growing header block
slow-readmemoryrequest a large response, read it ~1 B/s (read-side Slowloris)
keepalive-hoardconnectionsmany complete keep-alive requests held idle (idle-timeout / max-conn)
content-length-preallocmemoryhuge declared Content-Length, no body (CVE-2026-5440)
expect-100-continuememoryExpect: 100-continue + big CL, body withheld (CVE-2020-10705)
large-uploadmemorya huge uncompressed body delivered fast (whole-body buffering)
huge-chunkmemorychunked, one 0x7fff…-byte declared chunk (pre-alloc)
chunk-size-overflowcrashchunk-size near 2⁶⁴ → arithmetic overflow
chunk-extension-floodCPU1-byte chunks with kilobyte chunk-extensions (CVE-2023-39326)
tiny-chunk-floodCPUmillions of 1-byte chunks (per-chunk cost)
trailer-floodmemorychunked body + an unbounded trailer section (CVE-2026-22815)
header-bombmemoryone enormous header value (per-field / total size cap)
header-count-floodmemorytens of thousands of tiny header fields (CVE-2026-28980)
duplicate-header-mergeCPUthe same header name ×N → O(n²) join before dispatch
pipelining-floodmemorythousands of pipelined requests, responses unread (CVE-2026-54273)

HTTP/1.1 parser-differential

Run against the POST /parse endpoint. An entry that implements /parse parses the request body with its standard facilities (JSON, form/multipart, request decompression) — so these measure the framework's real parser limits. An entry that does not implement /parse (returns 404/405) shows 🔌 no-endpoint and is not scored (neither pass nor fail):

AttackExhaustsWhat it does
decompression-bombmemorya gzip body that inflates to GBs
decomp-codec-bypassmemorya deflate bomb — caps are often gzip-only (CVE-2026-42587)
hashdosCPUmany form keys → O(n²) in an un-seeded hash map
json-nesting-bombCPUa deeply nested JSON body → stack overflow
json-bignumCPUone gigantic JSON number → BigDecimal / quadratic int (CVE-2023-4043)
multipart-floodmemorymultipart/form-data with tens of thousands of parts
redos-headersCPUheader values crafted for catastrophic regex backtracking
qvalue-cpu-bombCPUa huge Accept-Language list → O(n²) q-value sort (CVE-2020-27223)
cookie-floodmemorymany cookie pairs under the header-byte cap (CVE-2025-58186)
percent-decode-quadraticCPUa path of many %XX tokens → super-linear decode (CVE-2026-45822)
range-bombmemorymany ranges on a static file → multipart/byteranges amplification (CVE-2011-3192)

HTTP/2

Only when the entry has an h2c profile (baseline-h2c/json-h2c), over the cleartext h2c listener on :8082. (The attacks target :8082 directly, so an entry that does h2 only over TLS isn't covered yet — gating it in would fire the attacks at an unbound port.)

AttackExhaustsWhat it does
continuation-floodmemoryendless CONTINUATION frames with no END_HEADERS (CVE-2024-27316)
hpack-bombmemorya ~4 KB HPACK entry referenced by thousands of 1-byte indices (CVE-2016-6581)
rapid-resetCPUopen-then-immediately-RST streams in a tight loop (CVE-2023-44487)
madeyouresetCPUforce server-side RST via WINDOW_UPDATE(0) (CVE-2025-8671)
settings-flood / ping-flood / empty-frame-floodCPUframe floods the peer must process (CVE-2019-9515/9512/9518)
window-stallmemorymany streams at a 0 flow-control window, never granted (CVE-2019-9517)

WebSocket

Only when the entry has a /ws profile (echo-ws/echo-ws-pipeline/echo-ws-limited), over /ws on :8080:

AttackExhaustsWhat it does
ws-giant-framememorya frame header declaring a 2⁶³ payload length (receive-buffer pre-alloc)
ws-fragment-bombmemoryendless continuation frames, no FIN (CVE-2026-48779)
ws-deflate-bombmemorya permessage-deflate frame inflating to GBs (CVE-2026-1526)
ws-pong-floodmemoryflood PINGs, never read → unbounded auto-PONG queue (CVE-2026-109932)

TLS

Only when the entry has a TLS listener (json-tls/static-tls/8gbit on :8081, else baseline-h2/static-h2 on :8443):

AttackExhaustsWhat it does
tls-handshake-floodCPUmany full handshakes → the server pays asymmetric crypto each time (THC-SSL-DOS class)
dheatCPUoffer only DHE → a modular exponentiation per handshake (CVE-2002-20001); reports immune if no DHE is offered

The /parse endpoint

The parser-differential attacks need a route that actually parses the request body — benchmark handlers like /echo just stream bytes back, so the attack never reaches a parser and would falsely read survived. An entry opts into these attacks by implementing:

POST /parse   — parse the request body with the framework's standard facilities:
                application/json      → the JSON parser (depth / number limits)
                x-www-form-urlencoded → the form parser (key count / hashing)
                multipart/form-data   → the multipart parser (part count)
                other (+ Content-Encoding) → request decompression + buffering
              Return 200 with anything (the body content is not checked).

If an entry doesn't implement /parse (returns 404/405), those attacks report 🔌 no-endpoint and drop out of its score — honest, rather than a phantom survived. aspnet-minimal implements it as the reference (System.Text.Json + ReadFormAsync + UseRequestDecompression).