CMSC818I: Advanced Topics in Computer Systems; Large Language Models, Security, and Privacy

Classroom: IRB 2107 Class hours: Monday and Wednesday, 3:30pm - 4:45pm

Instructor: Yizheng Chen Email: yzchen@umd.edu Office Hours: Monday 4:45 - 5:45pm, in IRB 5224

TA: Abhinav Rao Email: asura@umd.edu Office Hours: Wednesday 12pm - 1pm, in IRB 5112

Lectures

Date	Topic	Paper
08/31	Introduction	Syllabus
09/02	Background, Prompt Injection	AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents Defeating Prompt Injections by Design
09/07	Labor Day
09/09	LLM Memorization and Data Leakage	BenchChecker: Assessing the Credibility of Bug-Fixing Benchmarks for LLMs
09/14	Secure Code Generation	Give LLMs a Security Course: Securing Retrieval-Augmented Code Generation via Knowledge Injection
09/16	Secure Code Generation	Securing Retrieval-Augmented Code Generation via Contextual Knowledge Injection: A Case for Embedded IoT Applications
09/21	Privacy	Agent Tools Orchestration Leaks More: Dataset, Benchmark, and Mitigation
09/23	Privacy	Privacy Reasoning in Ambiguous Contexts
09/28	Agent Tool Use	Les Dissonances: Cross-Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents
09/30	Agent Tool Use	AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations
10/05	Package Hallucinations	We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs
10/07	Package Hallucinations	Evaluating Inference-Time Defenses Against Package Hallucination in LLM-Generated Code
10/12	Fall Break
10/14	Model Backdoor	Your Compiler is Backdooring Your Model: Understanding and Exploiting Compilation Inconsistency Vulnerabilities in Deep Learning Compilers
10/19	Vulnerability Detection	Patch-Guided Vulnerability Detection: Extracting Java API Security Rules via Attack–Defense Cross-Analysis
10/19	Midterm Project Report Due
10/21	RL for Fuzzing	Specializing Language Models for Textual Fuzzing via Reinforcement Learning
10/26	Vibe-Coded Applications	Understanding the (In)Security of Vibe-Coded Applications
10/28	Agentic Workflow	Comment and Control: Hijacking Agentic Workflows via Context-Grounded Evolution
11/02	Malicious agent skills	Cloak and Detonate: Scanner Evasion and Dynamic Detection of Agent Skill Malware
11/04	Malicious agent skills	No Attack Required: Semantic Fuzzing for Specification Violations in Agent Skills
11/09	Agent Social Network	WeClawArena: An Auditable Sandbox and Benchmark for Cross-User Agents Collaboration and Security in Human-Centered Agent Networks
11/11	Agent Social Network	“Humans welcome to observe”: A First Look at the Agent Social Network Moltbook
11/16	Evolving Coding Agents	Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents
11/18	Evolving Coding Agents	EVOMAL: Self-Poisoning in Self-Evolving Coding Agents
11/23	Thanksgiving break
11/25	Thanksgiving break
11/30	Project Presentation
12/02	Project Presentation
12/07	Project Presentation
12/09	Project Presentation
12/14	Final Project Report Due