The project
DebugABot is a deep tech research project dedicated to the governance of autonomous AI agents and embodied intelligent systems. The problem it addresses is concrete: as artificial intelligence systems acquire greater autonomy, it becomes essential to have tools to identify them, diagnose their behaviour and intervene when necessary.
The nine primitives
DebugABot’s architecture is structured around nine primitives organised into three operational phases:
Identify — recognise and trace:
- Model Fingerprinting: unique identification of the running AI model through its behavioural footprint.
- Blame Attribution: responsibility assignment for actions performed by an autonomous agent.
- Multi-Agent Tracing: tracking of interactions in complex multi-agent systems.
Diagnose — analyse and evaluate:
- Sycophancy/Deception Detector: detection of sycophantic or deceptive behaviours in models.
- Human Index Score: measurement of the degree of alignment between the agent’s behaviour and human standards.
- Active Ethical Injector: dynamic injection of ethical constraints into the agent’s decision-making process.
Intervene — control and contain:
- Kill Switch: immediate and safe shutdown of an autonomous agent.
- Behavioral Controller: real-time modulation of the agent’s behaviour.
- Rogue Intelligence Containment: containment of systems that deviate from expected behaviour.
The technologies
The research stack includes Transformers, Diffusion Models, State-Space Models, World Models, neuromorphic architectures, Embodied AI and Runtime Monitoring.