GHSA-mpqx-q82w-9x46: 
Python vulnerability analysis and mitigation

Summary

When a validation was aborted because the document contained more errors than allowed by max_errors (100 by default), validate() raised a single module-level ValidationAbortedError instance. Python appends the frames of every raise to the existing __traceback__ of an exception that is raised again, so every aborted validation extended the traceback of this shared error. That traceback kept the frames, and with them the document, the validation context and all errors of every aborted validation, alive for the lifetime of the process.

Details

Any client can trigger an aborted validation by sending a small document with more than 100 validation errors, e.g. 101 unknown fields. No authentication is required, and since validation runs before execution, depth limiters and complexity analyzers do not prevent it. Each such request permanently retained about 124 KiB on CPython 3.12, without an upper limit, so memory grew with every request until the process was restarted or killed. GraphQL.js uses a shared error object at the same place, but JavaScript errors capture their stack trace only once when they are created, so it is not affected. GraphQL-core 3.2 creates a new error for each aborted validation and is not affected either.

Proof of concept

import gc
import tracemalloc
from graphql import build_schema, parse, validate
schema = build_schema("type Query { a: Int }")
document = parse("{ " + " ".join(f"f{i}" for i in range(101)) + " }")
tracemalloc.start()
validate(schema, document)
gc.collect()
base = tracemalloc.get_traced_memory()[0]
for _ in range(200):
    validate(schema, document)
gc.collect()
print((tracemalloc.get_traced_memory()[0] - base) // 1024, "KiB")

With GraphQL-core 3.3.0 on CPython 3.12, the retained memory grows from 147 KiB after one call to about 25 MB after 200 calls. With the fix, it stays constant at about 180 KiB.

Impact

Denial of service through memory exhaustion. A few thousand small requests are enough to push a worker past a typical 256 MB or 512 MB container limit and trigger an OOM kill. No data is disclosed or modified.

Patches

Fixed in 3.3.1. validate() now creates a new ValidationAbortedError for every aborted validation and removes its traceback and exception context before returning it, so that the returned errors do not keep the frames of the validation alive. The module-level validation_aborted_error instance, which was not exported, has been removed.

Workarounds

If you cannot upgrade, clear the traceback of the shared error after each validation:

from graphql.validation import validate as validate_document
from graphql.validation.validate import validation_aborted_error
def validate(*args, **kwargs):
    try:
        return validate_document(*args, **kwargs)
    finally:
        validation_aborted_error.__traceback__ = None

Pass this function instead of the original validate where your framework allows it.

References

• https://github.com/graphql-python/graphql-core/issues/287

Credits

Reported by @riodw. @Fire162 submitted a first fix in #288.