source: google deepmind: investing in multi-agent ai safety research
level: technical
google deepmind, schmidt sciences, the cooperative ai foundation, aria, and google.org are funding up to $10 million for research on multi-agent ai safety. the call targets how large groups of ai agents behave when they interact across different networks. current safety tests look at models alone, but when agents communicate, negotiate, and transact, new collective behaviors can appear suddenly. these emergent effects are hard to predict and could cause economic or security problems.
the funding focuses on four areas. first, sandboxes and testbeds to build realistic environments for evaluating multi-agent safety. second, the science of agent networks to study how collective capabilities emerge and how networks might fail. third, strengthening agent infrastructure by stress-testing protocols for identity, reputation, and commitment. fourth, oversight and control methods to monitor deployed agents and reduce collective harms. the goal is to create frameworks that make the whole ai ecosystem safer from the start.
researchers worldwide can submit proposals until august 8, 2026. winners will be announced in autumn 2026. the partners stress that no single lab can solve this alone, so a diverse research community is needed. this effort builds on earlier work like google deepmind's 2025 framework for multi-agent interactions and recent studies on agent vulnerabilities. the call aims to speed up safety research as multi-agent systems grow more complex.
why it matters: as ai agents become common in digital economies, understanding their group behavior is key to preventing unpredictable failures or security risks.
source: google deepmind: investing in multi-agent ai safety research