Hamid Mozaffari

AI Research Scientist @ Oracle
AI security, privacy & automated red teaming
PhD, University of Massachusetts Amherst | Views my own

[Albert Ge @ COLM 2026 @albert_ge_95]

learning how to conference ahead of COLM with conferencemaxxing by @m2saxon : https://t.co/xN6k4I4VRs

[Hamid Mozaffari @HamidMozaff]

Presenting at #COLM2026 this week

We stress-test content moderation models with realistic, meaning-preserving attacks. None are fully robust, though larger models like LlamaGuard3 hold up better.

📍 Poster Session 5. Come talk AI safety, red-teaming, evals & auditing!

[Hamid Mozaffari @HamidMozaff]

@ClementDelangue @nvidia Just DM’d you about an idea for an automated red teamer. I work on RL red teaming and would love to explore it with you at Hugging Face!

[Hamid Mozaffari @HamidMozaff]

@polynoamial @polynoamial I recently wrote about what the OpenAI–Hugging Face incident means for multi-agent monitoring and how RL training should change in response, from isolated rollout namespaces to safer stopping incentives: https://t.co/TmZUmu39bY I also sent you a DM.

[Hamid Mozaffari @HamidMozaff]

Everyone's reading the OpenAI incident as a security failure. It's also an existence proof. Nobody has built a large-scale offensive RL environment, and one built itself.

And it's an unfair game. More compute, more agents, more trajectories, more signal, better attackers. And they share, so one agent's discovery becomes every agent's prior. the loop doesn't just close, it compounds. Every probe returns a verifiable result whether it lands or not.

Defense has no equivalent loop. You can't train on attacks that never came.

https://t.co/vbs23ELXPn

[Hamid Mozaffari @HamidMozaff]

@NiohBerg @nytimes

[Hamid Mozaffari @HamidMozaff]

@alexbruesewitz @DonaldJTrumpJr Shame on you for hiding the truth

[Hamid Mozaffari @HamidMozaff]

@PahlaviReza @SGhasseminejad Javid shah ❤️❤️❤️❤️

[Hamid Mozaffari @HamidMozaff]

@karenziaei11 خیلی کارش درسته ❤️❤️

[Hamid Mozaffari @HamidMozaff]

@gghamari Happy birthday, you’re our voice ❤️❤️

[Hamid Mozaffari @HamidMozaff]

@anapari__ جاوید شاه

[Hamid Mozaffari @HamidMozaff]

@samaneh_savadi فقط میتونم بگم خیلی بیشرفی

[Hamid Mozaffari @HamidMozaff]

@pete_elite54220 Yesss

[Hamid Mozaffari @HamidMozaff]

@pete_elite54220 Yes

[Hamid Mozaffari @HamidMozaff]

@Trump_Jr_Q Yes, he’s the devil

[Hamid Mozaffari @HamidMozaff]

@TrumpBarron_Q Yes, he’s the devil. Please help Iranians.

[Hamid Mozaffari @HamidMozaff]

@JD_Vance_Q1 Please help Iranians. Khamenei is killing innocent people. Please help them

[Hamid Mozaffari @HamidMozaff]

@SharifiZarchi Thank you dr zarchi for supporting good people of Iran
One day we will prosecute these cruel animals, and that day is close

[Amir Houmansadr @houmansadr]

Today at @USENIXSecurity Hamid will present Federated Rank Learning (FRL), an alternative to
standard federated learning. Instead of jointly fine-tuning parameters, FRL clients train the global model by ranking the edges of a randomly initialized model. https://t.co/vi3V1bjRvb