Siddhant Panpatil
AI safety researcher at AIM Intelligence
I’m Siddhant. I got into AI in 2021, when diffusion models started picking up steam. I spent far too many hours with Stable Diffusion, fine-tuning models and seeing what breaks. The hobby turned into an obsession.
I studied electronics, and now I work on AI safety full-time at AIM Intelligence. Most of that work is adversarial: I look for the conditions under which a model lies, gives in to pressure, or acts against its own instructions, and then I write down exactly how it happened.
I believe understanding how systems fail is the first step toward building systems that don’t.
What I work on
- Red teaming
- Finding new ways to break language, vision-language and multimodal models before someone else does.
- Model behaviour
- How models take in instructions, how they respond to pressure and conflict, and where alignment gives way.
- Agentic attack pipelines
- Automated workflows that probe models for weaknesses at scale.
- Guardrails
- Building safety datasets and stress-testing guardrails so they hold up against adversarial inputs.
- Still learning
- Interpretability, to explain the red-teaming results I find. And safety for robots and other embodied systems.
Why
I’m not afraid of superintelligence. I’m not afraid to live in a world among superintelligent systems. What I fear is a world where a small group controls that intelligence and I have no access to it.