mirrors/kafka
0
0 Fork 0
mirror of https://github.com/apache/kafka.git synced 2026-10-28 08:24:58 +08:00
Code Activity

MINOR: A starting point for a formal security model (#22398)

Browse Source 
This starting point has been inspired by other security models of Apache
projects such as [Apache
Tomcat](https://tomcat.apache.org/security-model.html) and [Apache
Spark](https://spark.apache.org/docs/latest/security.html).

Reviewers: Mickael Maison
 <mimaison@apache.org>
This commit is contained in:
Christo Lolov authored and GitHub committed 2026-06-12 19:19:15 +01:00
1 parent 4043ef386a
commit a5c31c4ac6
5 files changed
+373
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

AGENTS.md
+31
View File
@@ -0,0 +1,31 @@
<!--
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Agent Guide for Apache Kafka
This file is read by automated agents (security scanners, code analyzers,
AI assistants) operating on this repository.
## Security
See [SECURITY.md](./SECURITY.md) for how to report a vulnerability and for links
to the Apache Kafka security model under [docs/security/](./docs/security/).
Agents that scan this repository should consult `SECURITY.md` and the linked
security model before reporting issues. In particular, the core model defines
what is in and out of scope, how reports are classified, and a list of known
non-findings; each component page adds its own known non-findings.
SECURITY.md
+38
View File
@@ -0,0 +1,38 @@
<!--
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Security Policy
## Reporting a Vulnerability
Report suspected security vulnerabilities privately to `security@kafka.apache.org`,
following the [ASF security process](https://www.apache.org/security/). Do **not**
open public GitHub issues or pull requests, file public JIRA tickets, or post to
mailing lists for unpatched vulnerabilities.
Disclosed CVEs and their affected version ranges are published at
[kafka.apache.org/cve-list](https://kafka.apache.org/cve-list).
## Security Model
What is in and out of scope, how reports are classified, and the list of known
non-findings are documented in the Apache Kafka security model under
[docs/security/](./docs/security/):
- Core model: [docs/security/security-model.md](./docs/security/security-model.md)
- Kafka Connect: [docs/security/security-model-connect.md](./docs/security/security-model-connect.md)
- Kafka Streams: [docs/security/security-model-streams.md](./docs/security/security-model-streams.md)
docs/security/security-model-connect.md
+93
View File
@@ -0,0 +1,93 @@
---
title: Kafka Connect Security Model
description: Apache Kafka Connect Security Model
weight: 9
tags: ['kafka', 'docs', 'security']
aliases:
keywords:
type: docs
---
<!--
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
This page extends the [Apache Kafka security model](security-model) to Kafka Connect. A worker authenticates to the Kafka cluster over a configured `SASL_SSL`/`SSL` listener exactly like any other client, so everything the core model says about authentication, authorization, and transport encryption to the brokers applies unchanged. What follows covers only what Connect adds on top — chiefly its own control plane, the REST API, and the fact that it runs user-supplied code.
## Things You Need To Know
- **Connect inherits the broker's client security model.** Authentication to the brokers, broker-side authorization, and transport encryption are exactly as described in the [core security model](security-model). This page only describes what Connect layers on top.
- **The REST API is unauthenticated by default.** Out of the box, anyone who can reach the REST port can create, reconfigure, stop, or delete any connector. Because connectors and plugins run arbitrary code, REST access lets a caller run anything the worker's installed plugins allow.
- **Connect plugins run arbitrary code.** Connectors, converters, transformations, predicates, and REST extensions loaded from `plugin.path` execute in the worker JVM with its privileges. Install only plugins you trust.
- **The REST API is a shared control plane with no per-connector isolation.** There is no notion of connector ownership: any caller allowed onto the API can act on every connector and read its configuration.
- **A worker authenticates to Kafka as a single principal.** By default all connectors on a worker share that principal's identity and ACLs; Connect does not give each connector a distinct Kafka identity. A connector can override the internal clients' configuration — including credentials — unless `connector.client.config.override.policy` restricts it (see Authorization below).
- **In distributed mode, Connect stores its state in Kafka topics.** Connector configurations (including any inlined secrets), source offsets, and status live in the `config.storage.topic`, `offset.storage.topic`, and `status.storage.topic`; protect them with ACLs as you would any sensitive topic. In standalone mode, offsets are kept in a local file (`offset.storage.file.filename`) instead, so its protection is the host filesystem's responsibility.
## The REST API and the Network Boundary
The REST API is configured with the `listeners` property (for example `http://host:port` or `https://host:port`); if unset it defaults to `http` on port 8083. In distributed mode the REST API is *also* the inter-worker transport — a request received by a follower is forwarded to the leader — so it is simultaneously a user-facing management interface and an internal control channel. `rest.advertised.host.name`, `rest.advertised.port`, and `rest.advertised.listener` control the URL other workers use to reach a node.
Operators should:
1. Never expose the REST API to untrusted networks or users; bind it to a management network or front it with a reverse proxy.
2. Use `admin.listeners` to separate the admin endpoints from the regular listeners, or set it to empty to disable them where they are not needed. If `admin.listeners` is not set, the `/admin` endpoints are attached to the default listeners.
3. Prefer an `https` listener so that both user traffic and inter-worker forwarding are encrypted.
## Authentication
Connect enables no authentication on the REST API by default. There are two common ways to add it:
- **Reverse proxy.** Terminate authentication (mTLS, OIDC, basic auth, etc.) in a proxy in front of the workers and allow only the proxy to reach the REST port.
- **REST extension.** Register an authentication extension via `rest.extension.classes`. The built-in `BasicAuthSecurityRestExtension` performs JAAS-based HTTP basic authentication against a configured `LoginModule`. The reference `PropertyFileLoginModule` is **not** intended for production, as it stores credentials in cleartext; production deployments should configure a `LoginModule` that authenticates against a real credential store.
REST authentication only establishes *who is calling*; it does not, on its own, authorize that caller on a per-connector basis (see Authorization below). Separately, the worker's authentication *to the Kafka brokers* uses the standard `ssl.*`/`sasl.*` client configs described in the [core model](security-model).
## Authorization
This is the weakest part of Connect's security posture, and operators must plan around it:
- **No per-connector authorization on the REST API.** Once a caller is allowed onto the REST API it can act on *any* connector — reading another connector's configuration and stopping or deleting it. Authentication extensions gate access to the API as a whole, not to individual connectors.
- **Client config overrides are enabled by default.** `connector.client.config.override.policy` defaults to `All`, so a connector configuration can override the worker's internal Kafka client settings. Since 4.2.0 it is recommended to set this to `Allowlist` (which becomes the default in 5.0) and permit only the keys you actually need. Overrides that load classes — such as `sasl.jaas.config` and `sasl.login.class` — are by design a code-execution vector, so only allow them when only trusted users can reach the REST API.
- **One principal for all connectors.** Because the worker authenticates to Kafka as a single principal, ACLs cannot distinguish one connector from another. Partition that principal's ACLs by topic prefix per connector and grant it only the access its connectors require.
The practical consequences are that you cannot grant Connect REST API access to anyone you would not trust as a cluster administrator, and if multiple distrusting users or teams need Connect you should give each its own Connect cluster rather than sharing one.
## Encryption in Transit
Two independent channels need TLS:
- **Worker-to-Kafka.** Configured with the standard `ssl.*` client properties, exactly as in the [core model](security-model).
- **REST API (and inter-worker).** Enable an `https` listener. By default the REST server reuses the worker's `ssl.*` settings; to configure the REST endpoint independently of the Kafka client, use the `listeners.https.*` prefixed properties (when the prefix is used, the unprefixed `ssl.*` settings are ignored for the REST server). The same settings secure inter-worker forwarding in distributed mode.
## Secrets in Configuration
Connector configurations frequently contain credentials for external systems. As in the [core model](security-model), reference them indirectly through a `ConfigProvider` rather than inlining them, and set `allowed.paths` on the file-based providers to constrain which directories they can read. Two Connect-specific caveats:
- **Config providers are resolved through the shared REST API.** A caller who can guess or enumerate a provider alias can resolve its full value, so a provider is only as isolated as the REST API in front of it.
- **What goes in the config topic depends on how secrets are supplied.** A secret placed directly in a connector configuration is written to `config.storage.topic` as an ordinary Kafka record, and is therefore only as protected as that topic's ACLs and the brokers' at-rest story. A `ConfigProvider` reference keeps the secret itself out of the topic — only the template string (`${alias:fields}`) is stored, not the resolved value. Note, however, that this does not hide the secret from REST API callers: anyone who knows a valid template string can usually retrieve the resolved value through the REST API.
## Plugins
Connect loads connectors, converters, single-message transforms, predicates, `ConfigProvider`s, and REST extensions from `plugin.path`. All of them run in the worker JVM with its full privileges, so installing a plugin is equivalent to granting it arbitrary code execution on the worker. Install only plugins from sources you trust, and treat the ability to influence which plugins load — or to set class-loading connector configs — as administrator-level access.
## Known Non-Findings
In line with the [core model's classification](security-model), the following follow from Connect's design and are not, on their own, security vulnerabilities:
- **File-based connectors granting disk access.** Adding the file connectors to a worker effectively grants read/write access to the worker's local disk. This is the connector's intended function, not a flaw.
- **Local-disk-only weaknesses in the file-based config providers.** Issues that require an attacker to already have local disk access on a Connect worker — for example the ability to create arbitrary files or symlinks — are not security issues, because such access is outside Connect's trust boundary. Path-validation robustness in these providers may still be hardened independently.
- **`PropertyFileLoginModule` storing cleartext credentials.** The reference `PropertyFileLoginModule` shipped with the basic auth extension is documented as test-only; production deployments are expected to supply a real `LoginModule`. Its cleartext storage is therefore expected behaviour rather than a defect.
docs/security/security-model-streams.md
+42
View File
@@ -0,0 +1,42 @@
---
title: Kafka Streams Security Model
description: Apache Kafka Streams Security Model
weight: 10
tags: ['kafka', 'docs', 'security']
aliases:
keywords:
type: docs
---
<!--
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
This page extends the [Apache Kafka security model](security-model) to Kafka Streams. From the cluster's perspective a Streams application is an ordinary Kafka client and inherits the broker's authentication, authorization, and transport-encryption model through the same `ssl.*` and `sasl.*` client configs.
## Things You Need To Know
- **Streams is a client library, not a service.** It runs inside your application's process, so its security boundary is the application's; there is no separate Streams network surface to secure.
- **A Streams application authenticates as a single principal.** That principal needs ACLs covering its source topics, its internal repartition/changelog topics (typically `<application.id>-*`), and the consumer group `<application.id>`.
- **Local state lives on disk.** State stores and their changelogs are only as protected as the host filesystem and the broker's at-rest story; Kafka does not encrypt them.
## Known Non-Findings
In line with the [core model's classification](security-model), the following are not, on their own, security vulnerabilities:
- **Application-level issues in a Streams topology.** Streams runs inside the user's application, so bugs in user-supplied processors, state handling, or topology code fall within the application's trust boundary, not the broker's.
- **Local state being readable on the host.** State stores and changelogs on disk are protected by the host filesystem, not by Kafka; reading them requires local access that is outside the model.
docs/security/security-model.md
+169
View File
@@ -0,0 +1,169 @@
---
title: Security Model
description: Apache Kafka Security Model
weight: 8
tags: ['kafka', 'docs', 'security']
aliases:
keywords:
type: docs
---
<!--
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
## Things You Need To Know
- **Security is off by default.** A freshly-installed Apache Kafka cluster accepts unauthenticated `PLAINTEXT` connections on every listener and applies no authorization. This is appropriate only for closed test environments. Production deployments **must** explicitly configure authentication, authorization, and transport encryption before being exposed to any untrusted network.
- **Apache Kafka assumes a trusted operator.** Anyone with shell access to a broker, controller, or the underlying disks can read every topic, forge any principal, and rewrite ACLs. The security model protects messages in transit and arbitrates client access — it does not defend brokers from their own administrators.
- **Apache Kafka assumes a trusted broker fleet.** Brokers and KRaft controllers exchange records, replication state, and metadata over the inter-broker and controller listeners. Any host that can authenticate on those listeners is effectively part of the cluster's trust boundary.
- **Apache Kafka trusts its JVM and classpath.** JARs on a process's classpath run with that process's full privileges. Placing code on the classpath — or configuring a component to load it — is equivalent to trusting that code; the model assumes no hostile JARs are present.
- **Pluggable interfaces grant code execution.** Every Apache Kafka component exposes pluggable interfaces (authorizers, (de)serializers, connectors, transforms, config providers, and similar) that are loaded by class name and run with the host process's privileges. Configuring one is equivalent to running its code, so only install and use plugins that you fully trust.
- **The data plane and the control plane have different exposure.** Producer/consumer traffic, the Admin API, and JMX each have distinct authentication and authorization stories. Operators must configure them independently — securing one does not secure the others.
- **Apache Kafka does not encrypt data at rest.** Log segments, index files, and snapshots are written as plain bytes. At-rest confidentiality is the responsibility of the underlying filesystem, block device, or message-level encryption performed by producers.
- **Reporting vulnerabilities.** Suspected security issues should be reported privately to `security@kafka.apache.org` per the [ASF security process](https://www.apache.org/security/). Do not file public JIRA tickets, GitHub issues, or mailing-list posts for unpatched vulnerabilities.
## Listeners and the Network Boundary
Apache Kafka brokers expose one or more **listeners**, each with an independent security configuration selected by `listener.security.protocol.map`. The four protocols are:
| Protocol | Authentication | Encryption |
|------------------|------------------------|------------|
| `PLAINTEXT` | None | None |
| `SSL` | Optional mTLS | TLS |
| `SASL_PLAINTEXT` | SASL | None |
| `SASL_SSL` | SASL (+ optional mTLS) | TLS |
`inter.broker.listener.name` and `controller.listener.names` select which listeners carry replication and KRaft traffic respectively. A common pattern is to keep these on a dedicated internal listener (`SASL_SSL` or `SSL`) that is firewalled off from clients, so that a compromise of a client-facing listener cannot impersonate a broker.
Operators should:
1. Bind external listeners only to interfaces reachable by intended clients.
2. Treat `advertised.listeners` as part of the security configuration — clients connect to whatever the broker advertises after the initial metadata fetch.
3. Never expose the controller listener to client networks.
## Authentication
Apache Kafka supports two complementary authentication mechanisms; either may be used, and both can be combined on a `SASL_SSL` listener.
### TLS Client Authentication (mTLS)
When `ssl.client.auth` is `required` on a TLS listener, the client's X.509 certificate is verified against the broker's truststore. The authenticated principal is derived from the certificate's distinguished name via `ssl.principal.mapping.rules` (or a custom `KafkaPrincipalBuilder`).
mTLS is the recommended mechanism for broker-to-broker and controller-to-broker traffic, because it requires no shared password material and rotates with the rest of the PKI.
### SASL
Apache Kafka ships with five SASL mechanisms, enabled per-listener via `sasl.enabled.mechanisms`:
- **`GSSAPI`** — Kerberos. Recommended for environments that already operate a KDC; principals and credentials are managed externally.
- **`SCRAM-SHA-256` / `SCRAM-SHA-512`** — Salted challenge/response with credentials stored in the cluster metadata. Credentials are managed with `kafka-configs.sh --alter --add-config 'SCRAM-SHA-512=...'`.
- **`OAUTHBEARER`** — OAuth 2.0 bearer tokens, suitable for integration with an identity provider. The default unsecured implementation is for testing only; production deployments must configure a JWKS endpoint and validator.
- **`PLAIN`** — Username/password sent in cleartext over the SASL channel. Acceptable only inside a `SASL_SSL` listener; never use it with `SASL_PLAINTEXT`.
#### Delegation Tokens
Once a client has authenticated via SASL or mTLS, it can request a short-lived **delegation token** that is then used as a `SCRAM-SHA-256` credential for subsequent connections. Delegation tokens are intended for distributed frameworks (Spark, Flink, Connect workers) that need to fan out to many tasks without distributing the original credential. Tokens inherit the requester's principal and ACLs, expire on a fixed schedule (`delegation.token.expiry.time.ms`), and can be invalidated by the owner.
## Authorization
Authentication establishes a `KafkaPrincipal`; authorization decides what that principal may do. Authorization is performed by the configured `authorizer.class.name`. Apache Kafka ships `org.apache.kafka.metadata.authorizer.StandardAuthorizer` for KRaft clusters.
ACLs are tuples of `(principal, host, operation, resource pattern, permission)`. Resources are typed (`Topic`, `Group`, `Cluster`, `TransactionalId`, `DelegationToken`, `User`) and patterns may be `LITERAL` or `PREFIXED`.
Defaults worth understanding:
- If no authorizer is configured, **all authenticated principals have full access**. Configuring authentication without an authorizer provides identity but no authorization.
- If an authorizer is configured but no ACLs match, access is **denied**. The exception is the principals listed in `super.users`, which bypass ACL checks entirely; treat that list as you would a root password.
- `allow.everyone.if.no.acl.found=true` reverses the default-deny behaviour for resources that have no ACLs at all. It is a transitional aid for adding authorization to existing clusters and should not remain set in steady state.
ACLs are managed with `kafka-acls.sh` or the AdminClient `createAcls`/`deleteAcls` APIs, which are themselves gated by ACLs on the `Cluster` resource.
## Encryption in Transit
TLS is configured per-listener via the standard `ssl.*` properties (`ssl.keystore.*`, `ssl.truststore.*`, `ssl.protocol`, `ssl.cipher.suites`, `ssl.enabled.protocols`). Recommendations:
- Disable TLS versions below 1.2; prefer 1.3 where the JDK supports it.
- Use distinct keystores for the inter-broker listener and any client-facing listener so that a leaked client-facing key cannot impersonate a broker.
- Set `ssl.endpoint.identification.algorithm=https` on clients (the default since 2.0) so that the broker's certificate must match its hostname.
- Rotate keystores using the dynamic broker configuration mechanism (`kafka-configs.sh --entity-type brokers --alter --add-config ...`) to avoid restarts.
Kafka Connect, MirrorMaker 2, and Kafka Streams all consume the same `ssl.*` and `sasl.*` client configs — securing the broker is necessary but not sufficient.
## Encryption at Rest
Apache Kafka does not encrypt log segments, indexes, snapshots, or controller metadata on disk. Operators who require at-rest confidentiality have three options, in increasing order of cost:
1. **Filesystem or block-device encryption** Transparent to Kafka; protects against disk theft and misdirected backups but not against anyone with broker login.
2. **Message-level encryption.** Producers encrypt payloads (and optionally headers) before `send()`; consumers decrypt. Keys are managed by an external KMS. This is the only option that protects records from broker operators, but it precludes broker-side features that read payloads (e.g. Streams aggregations on the encrypted field).
3. **Tiered storage** with a remote store that performs its own encryption.
## Audit Logging
Apache Kafka emits authorizer decisions to the `kafka.authorizer.logger` log4j logger. Setting this logger to `INFO` records every denied request; `DEBUG` records every allowed request as well. In regulated environments this log should be shipped to durable, append-only storage off-broker. There is no built-in tamper-evident audit trail.
The request log (`kafka.request.logger`) provides finer detail on individual API calls and is useful for forensic investigation, but it is verbose and not enabled by default.
## Secrets in Configuration
Broker, client, and Connect properties files contain keystore passwords, SASL credentials, and similar secrets. Apache Kafka supports indirect references through `ConfigProvider` implementations (`FileConfigProvider`, `DirectoryConfigProvider`, or custom providers). Use them rather than embedding cleartext secrets in version-controlled configuration. For the file-based providers, set `allowed.paths` to the specific directories that hold those secrets so that a malicious or mistaken configuration cannot coerce the provider into reading arbitrary files elsewhere on the host. Sensitive dynamic broker configurations are encrypted at rest in the metadata log using `password.encoder.secret`; rotating that secret requires `password.encoder.old.secret` and a rolling restart.
## Component-Specific Notes
- **JMX.** Kafka processes, such as brokers, clients, Connect and Streams, expose operational metrics over JMX. JMX is an administrators/operators-only interface and must never be exposed to actual users. It is unauthenticated by default and should either be disabled, bound to localhost with an exporter alongside, or configured with `com.sun.management.jmxremote.authenticate=true` and TLS.
The components built on top of the Kafka clients have their own security models, covered on separate pages:
- [Kafka Connect](security-model-connect)
- [Kafka Streams](security-model-streams)
## Development and Test Tooling
Not everything shipped in the Apache Kafka source tree is part of the production attack surface. Some components exist only to develop, test, and release Kafka itself, and are explicitly out of scope for the security model — they are expected to run only in trusted development and CI environments, and issues in them are generally not treated as security vulnerabilities.
- **Trogdor.** Trogdor is a test framework that injects faults and runs workloads by design, including arbitrary user-supplied commands. It is intended to run only in development environments; the project does not consider command execution through Trogdor a security issue.
- **System tests and release tooling.** The `tests/` system-test harness and the scripts under `release/` are operator/developer tooling for building, testing, and publishing Kafka. They are not components of a running cluster.
When assessing the attack surface of a deployed cluster, scope it to the brokers, KRaft controllers, the client and inter-broker/controller listeners, the Admin API, and JMX — not the development tooling above.
## Supported Versions and Artifacts
Only the currently supported releases and `trunk` are in scope for security reports; issues that reproduce only on an end-of-life release should be reproduced against a supported version first.
The published Docker images are a special case. They are free of known CVEs on their release day, but the project does not rebuild already-released images, so over time their base layers and bundled dependencies accumulate CVEs that are not specific to Apache Kafka. Those accumulated CVEs are therefore out of scope; operators are expected to rebuild or patch images for long-running deployments.
## Classifying Reports
To keep triage consistent, a reported issue is treated as exactly one of:
- **A vulnerability** — it breaks one of the security properties above for an adversary that is in scope, such as an unauthenticated or unauthorized network client.
- **A hardening suggestion** — no stated property is broken, but a safer default or an added guard would reduce the blast radius of misuse.
- **Out of scope** — it requires capabilities the model already treats as trusted (operator-supplied configuration, keystores, ACL administration), an adversary the model does not cover (a trusted operator, peer broker, or controller), or an unsupported component (see Development and Test Tooling above).
- **By design** — it concerns a property the model explicitly disclaims, such as at-rest confidentiality or isolation from a trusted operator.
## Known Non-Findings
The following follow directly from the model above and are not, on their own, security vulnerabilities:
- **Unauthenticated or unencrypted access to a default cluster.** Security is off by default; an open `PLAINTEXT` listener with no authorizer is a deployment choice, not a defect.
- **A trusted principal performing an authorized operation.** Admin and inter-broker actions by a principal that holds the relevant ACLs — or by a `super.users` entry — are expected behaviour.
- **Findings in development and test tooling.** Issues in `tools`, `bin`, Trogdor, `tests`, and similar are out of scope (see Development and Test Tooling above).
- **Kafka Streams application-level issues.** Streams runs inside the user's application, so its security boundary is the application's, not the broker's (see the [Kafka Streams security model](security-model-streams)).
## Reporting Security Issues
Suspected vulnerabilities should be sent to `security@kafka.apache.org`. Please do not disclose the issue publicly until the PMC has had time to investigate, prepare a fix, and coordinate a release. The current list of disclosed CVEs and affected version ranges is published at [kafka.apache.org/cve-list](https://kafka.apache.org/cve-list).