Shir Rozenfeld
ML researcher working on AI Safety through mechanistic interpretability.
Data Scientist @ Microsoft Security · PhD Candidate @ BGU
About
I'm a Data Scientist at the Cyber Security GenAI Lab in the Microsoft Security CTO Office, and a PhD candidate at Ben-Gurion University, where I do Technical AI Safety research in the Offensive AI Research Labstrong>. Earlier, I led the development of GAVELstrong>, an open-source framework for detecting misaligned behaviors and intentions in LLMs through activation analysis (ICLR 2026; Black Hat USA 2026 Briefing). Before that, I led a data analysis team on national security missions.
Research interests
LLM safety · activation-analysis · mechanistic interpretability · AI agents · model behavior & intention detection · offensive AI
Education
- M.Sc. in Data Science, Ben-Gurion University
- B.Sc. in Computer Software Engineering, Ben-Gurion University
News
🎤 Giving a talk at Black Hat USA 2026 Briefings
Presenting an extension of our work at one of the world's top cybersecurity conferences, August 5–6 in Las Vegas.
🎤 Gave a talk at Heron Lightning Talks, TLV
A lightning talk on recent advancements in AI Safety for the AI Security community.
🌍 EA Global: London 2026
Attended EAG London to connect with the AI safety community.
🎉 Paper accepted at USENIX Security 2026
Love, Lies, and Language Models: Investigating AI's Role in Romance-Baiting Scams — arXiv
Research
LLM safety & activation monitoring
Real-time detection of unsafe or misaligned behavior by monitoring a model's internal activations — the line of work behind GAVEL.
Interpretability & mechanistic analysis
Understanding the internal computations of language models, from individual features up to circuits.
AI agents for interpretability
Using agent-tool workflows to automate hypothesis generation, experiment orchestration, and validation of model-internal circuits.
Model behavior & misalignment
Characterizing how models behave, when they fail, and how they can be misused — including LLM-driven social engineering.
Contact
I'm glad to talk about LLM safety, interpretability, and offensive AI. The best ways to reach me are email and LinkedIn.