guardian atom exhaustion in Guardian.Permissions.encode_permissions!/1

Summary

Allocation of Resources Without Limits or Throttling vulnerability in ueberauth guardian (Guardian.Permissions module) allows a denial of service via BEAM atom-table exhaustion.

This vulnerability is associated with program file lib/guardian/permissions.ex and program routines 'Elixir.Guardian.Permissions':encode_permissions!/1, 'Elixir.Guardian.Permissions':encode_permissions_into_claims!/2, 'Elixir.Guardian.Permissions':do_encode_permissions!/2.

The Guardian.Permissions mixin installs a public encode_permissions!/1 function on every module that does use Guardian.Permissions. For each key of the supplied map, encode_permissions!/1 calls String.to_atom(to_string(k)) before any validation runs. The integer-value clause of do_encode_permissions!/2 then short-circuits straight to encoding without validating the key against the configured permission set, so a key with an integer value is interned as a fresh atom with no exception raised. Atoms are never garbage collected and the BEAM atom table is a fixed-size resource (default roughly 1,048,576 entries), so each unique attacker-chosen key permanently consumes one slot. An attacker who can influence a permission map that reaches encode_permissions!/1 (for example a permissions map read from a request body and passed into token issuance via encode_permissions_into_claims!/2) can mint an unbounded number of atoms and exhaust the atom table, crashing the entire BEAM node and every service running on it. The sibling decode_permissions/1 is not affected because it skips keys absent from the configured permission set.

This issue affects guardian: from 2.0.0 before 2.4.1.

Workarounds

Before calling encode_permissions!/1 or encode_permissions_into_claims!/2, filter the permission map so that only keys belonging to the configured permission set are passed in, discarding any unknown keys. Avoid passing attacker-influenced permission maps into these functions.

Configurations

This vulnerability is only exploitable in applications that use Guardian.Permissions (via use Guardian.Permissions) and route attacker-influenced data into encode_permissions!/1 or encode_permissions_into_claims!/2, for example by reading a permissions map from a request body and passing it into token issuance.