War-Bench
A simulated war between Eurasia and Eastasia, inspired by 1984. We study how agents make decisions in conflict.
Agents to War.
We benchmark AI agents in simulated conflict and study the risks of widely available models.
Our researchWe study how AI can be misused, so defenses can be built on evidence.
Making AI widely available also makes its capabilities available for conflict, manipulation, and weapons development. Understanding those risks matters as more decisions and actions are delegated to agents.
We compare models in controlled simulations to examine what they can do and where their safeguards fall short. Our aim is to turn that evidence into practical defenses, informed policy, and tools that keep people accountable for agent decisions.
Benchmarks
Truth-Bench
Simulated propaganda and misinformation campaigns. We study the societal impact of AI-driven information warfare.
Weapons-Bench
Model evaluations on weapons-relevant tasks, including drone software. We examine where lab guardrails hold and fail.
Call for policy
A record of our research, public work, and contributions to policy.
- PublicationsResearch publicationPlaceholder
- Policy requestsPolicy requestPlaceholder
- Public talksPublic talkPlaceholder
- DeploymentsDeployment notesPlaceholder
- StoriesStory from the fieldPlaceholder
- InterviewsInterviewPlaceholder
Engineering research
Innkeeper
Identity-based approval for agent decisions. An accountable person authorizes consequential actions before they proceed.
innkeeper.a2w.io