Skip to main content
Menu
Advisories

Found while validating the fix for issue #10597 at main e353330c5c

GHSA-9CX8-QJP8-HCMRPublished Updated GitHub

Description

Found while validating the fix for issue #10597 at main e353330c5c. Reported privately per SECURITY.md.

Context first, because it matters for how you read this: #10597 itself was filed by us as a PUBLIC issue, which was our mistake. The log-amplification half it reported is genuinely fixed. What follows is a residual we are NOT adding to that public thread, and the public note on #10846 frames it only as a log-level change.

#10597 residual, deepened. The log-amplification half is genuinely fixed: 40 KB of junk now produces one -ERR Protocol error and a close, and the record count no longer scales with the input. What survives is a pre-authentication connection hold, measured on a real socket against the test server:

*1\r\n:1\r\n 8 bytes -> no reply at all, connection STILL OPEN after 4s *1\r\n:1\r\n repeated 2000 times 16000 bytes -> no reply at all, connection STILL OPEN after 4s *1\r\n*0\r\n 8 bytes -> no reply at all, connection STILL OPEN after 4s

The code path is executeCommand's first arm: an array whose first element is not a String is logged at FINE and the method returns without a reply and without setting shutdown, so the read loop goes straight back to parseNext. The maintainer names the :1 shape in the "still open" note of #10846 and tracks it as #10845, describing it as a log-level change. Three things make it worth more than that framing, and none of them are in the public issue:

  1. It is reachable with no credentials. parseNext and this arm both run before executeCommand reaches the AUTH check, exactly as the original report established for the two arms that were fixed.
  2. It defeats the idle timeout. arcadedb.network.socketTimeout (30 s) closes an idle pre-auth connection. Eight bytes every 29 s keeps a connection non-idle forever at a cost of roughly 0.3 bytes per second per connection, with the server holding one RedisNetworkExecutor thread per connection. The connection-count and thread-count ceiling is therefore reachable by an unauthenticated client at negligible bandwidth. The arms that were fixed all now close, so this is the only pre-auth shape left that holds a thread.
  3. The regression test cannot see it. Issue10597RedisMalformedInputTest.countingLogger counts only records of level >= Level.INFO, and the same commit moved this arm's record from SEVERE to FINE. So logRecordCountDoesNotDependOnTheNumberOfJunkBytes is structurally blind to this path, and a future change that re-floods it at FINE would not be caught either.

Suggested remedy, same shape as the rest of the fix: that arm raises RedisProtocolLimitException, so run()'s single handler owns the one reply and the close, and the test's counter counts every level rather than >= INFO. Nothing else in the parser then has a no-reply-no-close exit.

No other security-relevant finding in this set. For completeness, the one adjacent item I checked and am not reporting as a weakness: #10601's reportClusterTokenRefusal is reachable pre-authentication, but its 60 s gate is deliberately placed in front of both the log write and the password-file read, and the code comment explains why the gate is per node rather than per caller (a per-caller map would be state an unauthenticated caller could grow). That reasoning is correct.

Related advisories

This advisory and its CVE also cover the following advisories, fixed in the same release (26.11.1):

  • GHSA-hphw-285h-wx93: The unauthenticated Redis log amplification already public as issue 10597 also destroys the server log history, because the shipped handler is a bounded rotating one

Affected versions

References

GitHub · ArcadeData/arcadedb