Cheshire Cat StrayCat Session Object Isolation & Working Memory Management

Vulnerability Key Information Summary Vulnerability Overview Vulnerability Type: A session object (Session Object) is created for each HTTP request and WebSocket connection, and is exposed to plugins. Core Mechanism: The framework creates a Session instance for each request/connection, and plugins can access this object via . Primary Functions: - Provides (working memory) for storing and retrieving session state. - Provides to record model interactions. - Provides the method for sending messages over WebSocket. Impact Scope Affected Component: The class in . Affected Objects: All plugins that access session state via . Key Data Structures: - : A dictionary structure used for storing session state. - : Records the history of model interactions. - and : Used for report generation. Remediation Session Isolation: Ensure each HTTP request and WebSocket connection uses an independent Session instance to avoid state pollution. Memory Management: Manage session state using the and methods of to ensure data consistency. Message Sending: Use the method to send messages, ensuring they are correctly delivered via WebSocket. POC/Exploit Code ```python class StrayCat: """ The Stray cat goes around tools, hooks and endpoints, making troubles """ def __init__(self): self.user_data = AuthUser() self.user_id = user_data.name + "0000" self.user_data = user_data Get working memory from cache or create a new one self.working_memory = self.load_working_memory_from_cache() def _repr_(self): return f"StrayCat(user_id={self.user_id}, user_name={self.user_name})" def _send_ws_json(self, data: Any): Run the coroutine in the main event loop in the main thread and wait for the result app = CheshireCat().fastapi_app ws_manager = app.state.websocket_manager ws_connection = ws_manager.get_connection(self.user_id) if not ws_connection: log.debug(f"No websocket connection is open for user {self.user_id}") return asyncio.run_coroutine_threadsafe( ws_connection.send_json(data), app.state.event_loop, ).result() def _build_ai(self) -> MessageBag: """Build data structure for output (response and why and memories) e.g. TODO: these 3 lines are a mess, simplify episodic_report = [ dict([("type", "Episodic"), ("id", d[0]) for d in self.working_memory.episodic_memories ] declarative_report = [ dict([("type", "Declarative"), ("id", d[0]) for d in self.working_memory.declarative_memories ] procedural_report = [ dict([("type", "Procedural"), ("id", d[0]) for d in self.working_memory.procedural_memories ] Why this response? why = MessageBag( episodic_report=episodic_report, declarative_report=declarative_report, procedural_report=procedural_report, memory={ "episodic": episodic_report, "declarative": declarative_report, "procedural": procedural_report, }, model_interactions=self.working_memory.model_interactions, ) return why def load_working_memory_from_cache(self): """Load the working memory from the cache.""" self.working_memory = ( self.cache.get_value(f"{self.user_id}.working_memory") or WorkingMemory() ) self.update_working_memory_cache(self) """Update the working memory in the cache.""" cache_item = CacheItem(f"{self.user_id}.working_memory", self.working_memory, -1) self.cache.insert(updated_cache_item) def send_ws_message(self, content: str """Send a message via websocket. This method is useful for sending a message via websocket directly without passing through the LLM. In case there is no connection the message is skipped and a warning is logged. Parameters ---------- content : str The content of the message. msg_type : str The type of the message. Should be either "notification" (default), "chat", "chat_token" or "error" Examples -------- Send a notification via websocket >>> cat.send_ws_message("Hello, I'm a notification!") Send a chat message via websocket """

Goal Reached Thanks to every supporter — we hit 100%!

Cheshire Cat StrayCat Session Object Isolation & Working Memory Management

More from this source