Shiba AI
Shiba AI is a research laboratory conducting technical AI safety research across alignment, interpretability, and governance. The team develops evolutionary approaches to interpretable multi-agent coordination, builds mechanistic interpretability tools, and examines security and governance frameworks for AI deployment. Their work includes publications at ICML 2026 and ECML PKDD 2026, as well as seminars on AI-assisted vulnerability discovery.
- Artificial Intelligence
Funding
Founders
Founder details are not available yet.
Product
Problem
Advanced AI systems are becoming increasingly capable and autonomous, yet the technical methods for ensuring they remain aligned with human intentions, interpretable to oversight, and governable across multi-agent deployments lag behind. This gap creates risks of unintended behavior, security vulnerabilities, and a lack of transparency that hampers safe deployment in real-world applications.
Solution
Shiba AI is a research laboratory conducting technical AI safety research spanning alignment, interpretability, and governance. The team pursues evolutionary approaches to developing interpretable constitutions for multi-agent coordination, builds mechanistic interpretability tools to understand model reasoning, and examines governance frameworks for AI deployment. Their work integrates security research with AI safety, including organizing seminars on vulnerability discovery in the age of AI with leading security researchers. The lab also contributes to field building through collaborations with academic institutions and participation in programs like the Technical Alignment Research Accelerator (TARA).
Target Audience
The primary audience consists of academic institutions, AI safety researchers, and industry practitioners seeking technical methods for aligning, interpreting, and governing increasingly capable AI systems.
Features
- Research on evolutionary approaches for generating interpretable constitutions that coordinate multi-agent systems, with a paper available on arXiv
- Mechanistic interpretability research spanning model reasoning, multimodal AI, and bias detection across multiple team members
- Multi-agent coordination and game theory research combining economics and computer science perspectives
- Active engagement with frontier AI security topics, including hosting seminars on automated vulnerability discovery and AI-assisted exploit pipelines
- Publications accepted at major conferences including ICML 2026 and ECML PKDD 2026 BIAS Workshop
- Governance research addressing bias, persona stability, and LLM oversight structures