Introduction
A client that never authenticates could push a MongoDB replica set node into unbounded memory growth. The client only had to ask the node to wait for a point in time it will never reach. MongoDB shipped fixes across seven server branches on Nov 10, 2026. The fix gates an internal read option instead of bounding the underlying data structure, and operators need to understand that choice.
The CVE record rates CVE-2026-120828 as CVSS 3.1 7.5 High and CVSS 4.0 8.7 High. The two scores come from two CVSS versions and do not conflict. The weakness is CWE 770, allocation of resources without limits or throttling. The impact is availability only, with no confidentiality or integrity impact. The sources we reviewed name no threat actors and report no exploitation in the wild, and we could not verify CISA KEV membership. That is a gap in the evidence. It does not confirm that the issue has gone unexploited.
Technical information
Root cause
MongoDB's readConcern document can carry an afterOpTime field, which tells the server to hold the read until the node has reached a given optime. The fix ticket (SERVER-138119) describes afterOpTime as an internal, deprecated option that was never meant for regular clients. Before the fix, the server did not stop ordinary clients from sending it.
To honor the option, the node creates internal wait records for the request. According to the CVE description, when the requested point in time is one the node will never reach, these records are:
- not limited in number,
- not reclaimed, and
- still present after the originating requests have ended.
Each such request leaves state behind, and nothing caps the total. This is a textbook CWE 770 condition: a client can make the server allocate memory without any bound.
Attack prerequisites
The CVSS 3.1 vector in the CVE record describes these conditions:
- attack vector: network
- attack complexity: low
- privileges required: none
- user interaction: none
- availability impact: high, with no confidentiality or integrity impact
The CVSS 4.0 vector likewise indicates low attack complexity, no privileges and no user interaction.
Attack flow
Based on the CVE description and the regression test shipped with the fix, the attack path looks like this:
- The attacker reaches a replica set node over the network. No credentials are needed.
- The attacker sends a command whose
readConcerncontainsafterOpTimeset to a point in time the node will never reach. The regression test shows that commonly available commands such asping,find,aggregate,insert,helloandexplaincan carry the field. - The node registers internal wait records for the request.
- The request ends, but the wait records are not reclaimed.
- The attacker repeats the request. The records accumulate without bound until the server runs out of memory and becomes unavailable.
We found no public proof of concept in the sources we reviewed.
Patch information
MongoDB fixed the issue under Jira ticket SERVER-138119, titled "Allow only system users to use readConcern afterOpTime". The ticket is resolved as Fixed, and the patch shipped on Nov 10, 2026 in these first fixed releases:
- 9.0 branch: 9.0.7
- 8.3 branch: 8.3.16
- 8.0 branch: 8.0.37
- 7.0 branch: 7.0.48
- 6.0 branch: 6.0.30
- 5.0 branch: 5.0.35
- 4.4 branch: 4.4.32
The 8.0 and 9.0 release notes list CVE-2026-120828 and SERVER-138119 among their fixes. MongoDB did not publish 8.0.37 binaries for RedHat/CentOS 8.3 and 9 on s390x. On those platforms, use 8.0.38 instead.
How the fix works: block the input instead of limiting the waiters
MongoDB could have capped or reclaimed the wait records. It chose instead to stop untrusted clients from sending the option at all. The 6.0 backport commit c4ecf98 ("Allow only internal users to use readConcern.afterOpTime (#65520)") adds checkAuthForAfterOpTime in src/mongo/db/commands.cpp.
The function opens with a cheap early exit. If the request has no readConcern object, or the object has no afterOpTime field, it returns at once, so normal traffic pays almost nothing:
void checkAuthForAfterOpTime(OperationContext* opCtx, const OpMsgRequest& request) { const auto readConcernElem = request.body[repl::ReadConcernArgs::kReadConcernFieldName]; if (MONGO_likely(readConcernElem.type() != BSONType::Object || !readConcernElem.Obj().hasField(repl::ReadConcernArgs::kAfterOpTimeFieldName))) return;
If afterOpTime is present, the caller must be authorized for ActionType::internal on the cluster resource. In practice, that means identities used inside the cluster or __system identities. The commit's test file shows that even a root user is rejected. Any other caller gets a hard failure before the command runs, so the server never registers a waiter:
uasserted(ErrorCodes::Unauthorized, "Not authorized to run command with readConcern.afterOpTime");
When the parameter described below is enabled, or the caller is internal, the server accepts and counts the request. New metrics record every outcome.
Where the check runs
The function is called from CommandInvocation::checkAuthorization, ahead of each command's own doCheckAuthorization:
} else { try { + checkAuthForAfterOpTime(opCtx, request); doCheckAuthorization(opCtx);
Because the check sits on a generic path rather than in individual commands, it closes the unauthenticated attack path described in the CVE. The regression test, jstests/auth/after_op_time_requires_internal.js, confirms the following:
- An unauthenticated connection sending
{ping: 1, readConcern: {afterOpTime}}getsUnauthorized. A plainpingstill works. find,aggregate,insert,ping,helloandexplainare rejected when they carryafterOpTime, and they run normally without it.- A
readConcernnested inside anexplainis also caught, because the inner command is authorized recursively. - The check runs on mongos as well as mongod. The router rejects the request before forwarding it to a shard.
There is one caveat. The hunk places the call in the else branch that follows a "blanket authorization" case, and the diff does not show what condition triggers that blanket branch. Check whether the new check applies under your own authentication configuration.
A temporary escape hatch, off by default
The commit adds a server parameter in repl_server_parameters.idl for deployments that still depend on the option:
- Name:
allowDeprecatedAfterOpTimeReadConcern - Default:
false - Can be set at startup or at runtime
- Described in the IDL as "only as a temporary escape hatch"
Setting it to true makes the server accept the option from every client again, including unauthenticated ones, and the vulnerable behavior returns. The test turns the parameter on, confirms that the previously rejected commands now succeed, and then turns it off again. Leave it at the default.
New metrics to see who still uses it
The patch adds a readConcernCounters.afterOpTime section to serverStatus, with accepted and rejected counters. These come from AfterOpTimeOps in read_concern_stats.idl and are populated by ServerReadConcernMetrics. According to the code comment, the counters let operators see how much the deprecated option is still used before MongoDB removes it. A rising rejected count after upgrading means something, legitimate or hostile, is still sending afterOpTime.
What the patch does not change
In the 6.0 backport diff, the only changes outside tests are the authorization gate, the parameter, the metrics and build wiring. Nothing caps or reclaims waiter records in the replication layer, and internal or __system principals can still create them. The fix removes the unauthenticated, network-reachable path to resource exhaustion. It does not bound the data structure itself.
The CVE and Jira records we reviewed document no interim workaround or separate configuration change for systems that cannot be upgraded yet.
Affected systems and versions
The affected product is MongoDB Server, in replica set deployments. The published affected ranges are:
- 9.0.0 up to but not including 9.0.7
- 8.3.0 up to but not including 8.3.16
- 8.0.0 up to but not including 8.0.37
- 7.0.0 up to but not including 7.0.48
- 6.0.0 up to but not including 6.0.30
- 5.0.0 up to but not including 5.0.35
- 4.4.0 up to but not including 4.4.32
For 8.0 on RedHat/CentOS 8.3 and 9 on s390x, the first available fixed build is 8.0.38. On fixed versions, any deployment that sets allowDeprecatedAfterOpTimeReadConcern to true accepts afterOpTime from all clients again and should be treated as exposed.
Vendor security history
The MongoDB Community Hub announcement for the Nov 10, 2026 releases is titled as an important security update requiring action. HKCERT published a bulletin covering multiple MongoDB vulnerabilities in the same cycle. MongoDB's published disclosure process includes a HackerOne bug bounty program, validation of submissions by its security team, and notification steps for key users, customers and partners. Its fiscal 2026 Form 10K says its information security program is designed to align with the NIST Cybersecurity Framework. These sources describe processes. They are not a complete record of past incidents.
References
- NVD entry for CVE-2026-120828
- CVE Program record for CVE-2026-120828
- MongoDB Jira SERVER-138119
- MongoDB commit c4ecf98 on GitHub
- MongoDB commit c4ecf98 via GitHub API
- MongoDB 8.0 release notes
- MongoDB 9.0 release notes
- MongoDB Community Hub: Important MongoDB security update available, November 10, 2026
- HKCERT: MongoDB Multiple Vulnerabilities
- MongoDB: Addressing Security Vulnerabilities
- MongoDB, Inc. fiscal 2026 Form 10K
- CISA Known Exploited Vulnerabilities Catalog



