OASIS: Multi-Agent AI Safety Lab

Advanced AI systems are increasingly likely to operate through interactions among multiple agents, tools, memory systems, and human decision-makers rather than as isolated models. This creates failure modes that single-model evaluations may miss, including hidden coordination, diffusion of responsibility, memory-driven drift, escalating autonomy, and loss of effective human oversight.

OASIS reduces existential AI risk by making these dynamics observable and testable before they appear in higher-stakes deployments. The project will build a reproducible multi-agent sandbox and compare governance mechanisms such as persistent audit logs, independent cross-checks, bounded authority, transparent memory, human veto, and preservation of agent histories across experiments.

The immediate output will be empirical evidence, evaluation protocols, and open-source infrastructure showing which mechanisms make dangerous coordination easier to detect, understand, and interrupt while preserving useful capabilities. These results can inform researchers, developers, evaluators, and funders working on safer agentic systems.

The longer-term impact is to contribute practical safety standards for multi-agent AI, reducing the probability that increasingly capable systems become opaque, unaccountable, or collectively uncontrollable.