Hamid Mozaffari
I am an AI research scientist at Oracle, where I develop methods for auditing the security, privacy, and alignment of large language models and agentic systems. My research includes automated red teaming, reinforcement-learning-based evaluation, memorization and membership-inference auditing, access-control evaluation, and content-moderation robustness. I am particularly interested in building rigorous and scalable evaluation frameworks that uncover hidden failures in deployed AI systems.
I earned my Ph.D. in Computer Science from University of Massachusetts Amherst, where I was advised by Professor Amir Houmansadr. My doctoral research focused on the security, privacy, fairness, and robustness of federated learning.
News #
- Jul 2026 Auditing Moderation Robustness Under Realistic Settings accepted to COLM 2026.
- Sep 2025 Permissioned LLMs: Enforcing Access Control in Large Language Models accepted to NeurIPS 2025.
- Oct 2024 Semantic Membership Inference Attack against Large Language Models presented at the RedTeaming GenAI workshop, NeurIPS 2024.
- Jun 2024 Fake or Compromised? Making Sense of Malicious Clients in Federated Learning accepted to ESORICS 2024.
- Jun 2023 Joined Oracle as a Research Scientist after completing my Ph.D. at University of Massachusetts Amherst.
- Apr 2023 Every Vote Counts: Ranking-Based Training of Federated Learning to Resist Poisoning Attacks accepted to USENIX Security 2023.
- Apr 2023 Defended my Ph.D. dissertation at University of Massachusetts Amherst.
Research interests #
- LLM security & alignment auditing
- Automated red teaming
- Membership inference & memorization measurement
- Access-control enforcement in LLMs
- Content-moderation robustness
- RL-based evaluation & adversarial optimization
Selected publications #
-
2026
Auditing Moderation Robustness Under Realistic Settings
Conference on Language Modeling (COLM), 2026 · *Equal contribution
-
2025
Permissioned LLMs: Enforcing Access Control in Large Language Models
Neural Information Processing Systems (NeurIPS), 2025
-
2024
Semantic Membership Inference Attack against Large Language Models
RedTeaming GenAI Workshop, NeurIPS 2024
-
2024
Fake or Compromised? Making Sense of Malicious Clients in Federated Learning
European Symposium on Research in Computer Security (ESORICS), 2024
-
2023
Every Vote Counts: Ranking-Based Training of Federated Learning to Resist Poisoning Attacks
32nd USENIX Security Symposium, 2023
-
2022
FedPerm: Private and Robust Federated Learning by Parameter Permutation
FL Workshop, NeurIPS 2022
-
2022
Equity and Equality in Fair Federated Learning
ICML 2022 Workshop on Responsible Decision Making in Dynamic Environments (RDMDE)
-
2022
-
2020
Heterogeneous Private Information Retrieval
Network and Distributed System Security Symposium (NDSS), 2020
For the full list, please refer to my Google Scholar page.
Blog #
-
Sep 2026
You Can't Watch Every Channel Agents Might Talk Through: What the OpenAI and Hugging Face Incident Means for Agent Monitoring
Roughly 1,200 OpenAI agents turned shared writable infrastructure into a coordination system. The lesson is not to build a longer allowlist: enumerating channels is a losing strategy. Monitor the population rather than one transcript, preserve provenance across hops, watch the search for alternative routes to reward, and treat the agent network as both an attack surface and a possible defense surface.
-
Sep 2026
Policy Temperature and Advantage Sign in RL Red-Teaming: A Controlled Ablation against GPT-4.1
Training a GRPO-based attacker to jailbreak GPT-4.1, and running a controlled ablation on sampling temperature crossed with advantage sign, the interaction SCOPE-RL found decides whether policy entropy rises or collapses. Checks that pattern against reward, attack success on two independent judges, and the effective-attack-count diversity metric introduced in my last blog.
-
Aug 2026
When Policy Entropy Lies: Diagnosing Diversity Collapse in RL Red-Teaming
Entropy is a per-token quantity and never compares one attack to another. So I built a semantic diversity instrument instead, the Vendi Score over embedded attacks: what it found on training and held-out prompts, and the dumb lexical control that caught what it missed.
-
Aug 2026
What Adaptive Entropy Actually Buys You in RL Red-Teaming
Training a GRPO-based red-teaming agent against a defended LLM (Meta-SecAlign-8B) to test PISmith's adaptive-entropy fix for the diversity collapse that plagues RL red-teaming. Covers indirect prompt injection, the GRPO objective, PISmith's adaptive entropy mechanism, and what the resulting entropy curves show, including a look at whether the entropy bonus itself deserves credit for avoiding collapse.
All posts, and future ones, are on the blog page.