AI Lab
Ghost in the Neural Shell
Anthropic showed tiny poisoned samples can implant backdoors in LLMs. I replicated this with GPT-2 using 50 poisoned examples, proving how easily training data tampering can compromise a model.
Mechanistic Interpretability
Mechanistic Interpretability
The Geometry of Refusal: Exploring LLM Activation Space Before and After Abliteration
What happens inside a neural network when it decides to refuse answering user’s prompt? We launched an expedition into the activation space to find out.
Agentic AI
The Domain-Specific Abliteration Paradox
How abliterating cybersecurity refusal collapsed nearly every safety domain — despite near-zero vector similarity except for the one that shared the most overlap