Active Proposals

A comprehensive testing suite for LLM adversarial attacks, building on the initial ABC grant #42. Focus on automated red-teaming pipelines for open source models.

Applying recent interpretability breakthroughs in SAEs to computer vision models to detect deception in image generation. Seeking seed funding for initial experiments.

Defining a new standard for measuring "loss of control" in multi-agent RL environments. Metadata standard proposal for universal adoption.