START WITH THE PROBLEM
The challenges shaping
AI safety.
Explore the problems researchers, engineers, policymakers, and organizers are trying to solve—then see who is already working on each one.
A WORKING MAP
Begin with what concerns you.
These categories overlap, and the field is still evolving. Use them as entry points—not as a final taxonomy.
Evaluating advanced AI
Build credible tests for dangerous capabilities, strategic behavior, autonomy, and safeguards before systems are widely deployed.
THE CENTRAL QUESTIONHow can we tell when an AI system has crossed a meaningful risk threshold?
Understanding how models work
Develop methods that reveal what models represent, how they reason, and when their internal processes differ from their stated behavior.
THE CENTRAL QUESTIONCan we inspect advanced systems well enough to catch deception or dangerous reasoning?
Controlling autonomous systems
Keep increasingly capable agents within legitimate authority, even when they operate for long periods, use tools, or coordinate with other agents.
THE CENTRAL QUESTIONHow do we preserve meaningful human control as agents become more capable and independent?
ORGANIZATIONS WORKING ON IT
AI security and misuse
Prevent model theft, malicious use, compromised agents, and cascading failures across the infrastructure on which advanced AI depends.
THE CENTRAL QUESTIONWhich defenses still work when both attackers and defenders can use capable AI agents?
Scalable oversight
Help people reliably evaluate work that is too fast, complex, or specialized for unaided human reviewers to check directly.
THE CENTRAL QUESTIONHow can humans supervise systems whose outputs exceed our own ability to verify them?
Governance and safety standards
Turn evidence about AI risk into enforceable rules, measurable standards, institutional capacity, and accountable deployment decisions.
THE CENTRAL QUESTIONWhich rules and institutions can keep pace with rapidly advancing capabilities?
Multi-agent coordination
Understand cooperation, conflict, bargaining, and emergent behavior when many AI agents and human institutions interact.
THE CENTRAL QUESTIONHow can independently developed agents cooperate without collusion, conflict, or systemic failure?
ORGANIZATIONS WORKING ON IT
International coordination
Create shared testing practices, risk thresholds, and response capacity across countries without reducing safety to the weakest consensus.
THE CENTRAL QUESTIONWhat can governments coordinate on before the highest-risk systems are globally deployed?
TURN INTEREST INTO ACTION
Find people and work connected to a challenge you care about.
Browse current roles and projects—or join a challenge circle to meet people coordinating around the same problem.