Python: Track instance attributes through type tracking

python/ql/lib/change-notes/2026-06-11-fix-type-tracking-instance-attributes.md

@@ -0,0 +1,4 @@
+---
+category: minorAnalysis
+---
+* Python type tracking now follows values stored in instance attributes such as `self.attr` across instance methods on the same class. As a result, analysis is more likely to recognize user-defined objects that are stored on `self` and used later in other methods, which may produce additional results.

python/ql/lib/semmle/python/dataflow/new/internal/TypeTrackingImpl.qll

@@ -167,11 +167,15 @@ module TypeTrackingInput implements Shared::TypeTrackingInput<Location> {
   }
 
   /** Holds if there is a level step from `nodeFrom` to `nodeTo`, which may depend on the call graph. */
-  predicate levelStepCall(Node nodeFrom, LocalSourceNode nodeTo) { none() }
+  predicate levelStepCall(Node nodeFrom, LocalSourceNode nodeTo) {
+    instanceFieldStep(nodeFrom, nodeTo)
+  }
 
   /** Holds if there is a level step from `nodeFrom` to `nodeTo`, which does not depend on the call graph. */
   predicate levelStepNoCall(Node nodeFrom, LocalSourceNode nodeTo) {
     TypeTrackerSummaryFlow::levelStepNoCall(nodeFrom, nodeTo)
+    or
+    localFieldStep(nodeFrom, nodeTo)
   }
 
   /**
@@ -317,6 +321,87 @@ module TypeTrackingInput implements Shared::TypeTrackingInput<Location> {
     )
   }
 
+  /**
+   * Holds if `ref` accesses attribute `attr` of `self`, where `self` is the first
+   * parameter of an instance method of `cls` (i.e. an access of the form `self.attr`).
+   *
+   * Static methods and class methods are excluded, since their first parameter is not a
+   * `self` instance reference.
+   */
+  private predicate selfAttrRef(Class cls, string attr, DataFlowPublic::AttrRef ref) {
+    exists(Function method, Name selfUse |
+      method = cls.getAMethod() and
+      not DataFlowDispatch::isStaticmethod(method) and
+      not DataFlowDispatch::isClassmethod(method) and
+      selfUse.getVariable() = method.getArg(0).(Name).getVariable() and
+      ref.getObject().asCfgNode().getNode() = selfUse and
+      ref.mayHaveAttributeName(attr)
+    )
+  }
+
+  /**
+   * Holds if `read` reads attribute `attr` from an instance of `cls`, where the instance
+   * is referred to from outside the methods of `cls` (i.e. an access of the form
+   * `instance.attr`, where `instance` is a reference to an instance of `cls`).
+   *
+   * This complements `selfAttrRef`, which only handles `self.attr` accesses inside the
+   * methods of `cls`. Unlike `selfAttrRef`, this depends on the call graph (via
+   * `classInstanceTracker`), so steps using it must be reported as `levelStepCall`.
+   */
+  private predicate instanceAttrRead(Class cls, string attr, DataFlowPublic::AttrRead read) {
+    read.getObject() = DataFlowDispatch::classInstanceTracker(cls) and
+    read.mayHaveAttributeName(attr)
+  }
+
+  /**
+   * Holds if `nodeFrom` is written to attribute `self.attr` in some instance method of a
+   * class, and `nodeTo` reads attribute `self.attr` in some (possibly different) instance
+   * method of the same class.
+   *
+   * This models flow through instance attributes (`self.foo`): a value stored into
+   * `self.foo` in one method can be read from `self.foo` in another method. Type-tracking
+   * handles the store and read steps via `AttrWrite`/`AttrRead`, but on its own it cannot
+   * relate the `self` of the writing method to the `self` of the reading method. Following
+   * the approach used for Ruby and JavaScript, we model this directly as a level step from
+   * the written value to the read reference, for any pair of methods on the class (not
+   * just from `__init__`).
+   *
+   * This is an over-approximation: it is instance-insensitive (it does not distinguish
+   * between different instances of the same class) and order-insensitive (it does not
+   * require the write to happen before the read), matching the precision of
+   * instance-attribute handling for Ruby and JavaScript.
+   */
+  private predicate localFieldStep(Node nodeFrom, LocalSourceNode nodeTo) {
+    exists(Class cls, string attr, DataFlowPublic::AttrWrite write, DataFlowPublic::AttrRead read |
+      selfAttrRef(cls, attr, write) and
+      nodeFrom = write.getValue() and
+      selfAttrRef(cls, attr, read) and
+      nodeTo = read
+    )
+  }
+
+  /**
+   * Holds if `nodeFrom` is written to attribute `self.attr` in some instance method of a
+   * class, and `nodeTo` reads attribute `attr` from an instance of the same class outside
+   * its methods (e.g. `instance.attr`).
+   *
+   * This is the cross-instance counterpart of `localFieldStep`: it relates a write of
+   * `self.attr` inside the class to a read of `attr` on a reference to an instance of the
+   * class. Identifying instances relies on the call graph (via `classInstanceTracker`), so
+   * this step is reported as `levelStepCall` rather than `levelStepNoCall`.
+   *
+   * Like `localFieldStep`, this is an over-approximation: it is both instance-insensitive
+   * and order-insensitive.
+   */
+  private predicate instanceFieldStep(Node nodeFrom, LocalSourceNode nodeTo) {
+    exists(Class cls, string attr, DataFlowPublic::AttrWrite write, DataFlowPublic::AttrRead read |
+      selfAttrRef(cls, attr, write) and
+      nodeFrom = write.getValue() and
+      instanceAttrRead(cls, attr, read) and
+      nodeTo = read
+    )
+  }
+
   /**
    * Holds if data can flow from `node1` to `node2` in a way that discards call contexts.
    */

python/ql/test/experimental/library-tests/CallGraph/InlineCallGraphTest.expected

@@ -1,9 +1,6 @@
 testFailures
 debug_callableNotUnique
 pointsTo_found_typeTracker_notFound
-| code/class_attr_assign.py:10:9:10:27 | ControlFlowNode for Attribute() | my_func |
-| code/class_attr_assign.py:11:9:11:25 | ControlFlowNode for Attribute() | my_func |
-| code/class_attr_assign.py:26:9:26:25 | ControlFlowNode for Attribute() | DummyObject.method |
 | code/class_super.py:50:1:50:6 | ControlFlowNode for Attribute() | outside_def |
 | code/conditional_in_argument.py:18:5:18:11 | ControlFlowNode for Attribute() | X.bar |
 | code/func_defined_outside_class.py:21:1:21:11 | ControlFlowNode for Attribute() | A.foo |

python/ql/test/experimental/library-tests/CallGraph/code/class_attr_assign.py

@@ -7,8 +7,8 @@ def __init__(self, func):
         self.direct_ref = my_func
 
     def later(self):
-        self.indirect_ref() # $ pt=my_func MISSING: tt=my_func
-        self.direct_ref() # $ pt=my_func MISSING: tt=my_func
+        self.indirect_ref() # $ pt=my_func tt=my_func
+        self.direct_ref() # $ pt=my_func tt=my_func
 
 foo = Foo(my_func) # $ tt=Foo.__init__
 foo.later() # $ pt,tt=Foo.later
@@ -23,7 +23,7 @@ def __init__(self):
         self.obj = DummyObject()
 
     def later(self):
-        self.obj.method() # $ pt=DummyObject.method MISSING: tt=DummyObject.method
+        self.obj.method() # $ pt=DummyObject.method tt=DummyObject.method
 
 
 bar = Bar(my_func) # $ tt=Bar.__init__

python/ql/test/library-tests/dataflow/typetracking/attribute_tests.py

@@ -151,13 +151,13 @@ def __init__(self): # $ tracked=foo
         self.foo = tracked # $ tracked=foo tracked
 
     def print_foo(self): # $ MISSING: tracked=foo
-        print(self.foo) # $ MISSING: tracked=foo tracked
+        print(self.foo) # $ tracked MISSING: tracked=foo
 
     def possibly_uncalled_method(self): # $ MISSING: tracked=foo
-        print(self.foo) # $ MISSING: tracked=foo tracked
+        print(self.foo) # $ tracked MISSING: tracked=foo
 
 instance = MyClass2()
-print(instance.foo) # $ MISSING: tracked=foo tracked
+print(instance.foo) # $ tracked MISSING: tracked=foo
 instance.print_foo() # $ MISSING: tracked=foo
 
 

python/ql/test/library-tests/frameworks/hdbcli/pep249.py

@@ -7,3 +7,30 @@
 cursor.executemany("some sql", (42,))  # $ getSql="some sql"
 
 cursor.close()
+
+
+# Connection stored in a class attribute (`self._conn`) and used in another method.
+#
+# This is detected because type tracking includes a level step modelling flow through
+# instance attributes: a value written to `self._conn` in one method (here `__init__`) can
+# be read back from `self._conn` (directly or via a getter) in any other method on the same
+# class. This follows the same approach used for instance fields in Ruby and JavaScript.
+class Database:
+    def __init__(self):
+        self._conn = dbapi.connect(address="hostname", port=300, user="username")
+
+    def get_connection(self):
+        return self._conn
+
+    def run_via_getter(self):
+        conn = self.get_connection()
+        cursor = conn.cursor()
+        cursor.execute("getter sql")  # $ getSql="getter sql"
+
+    def run_direct(self):
+        self._conn.execute("direct sql")  # $ getSql="direct sql"
+
+
+db = Database()
+db.run_via_getter()
+db.run_direct()