Google DeepMind has pledged $10 million to seed research on ecosystem safety. Funding comes from a coalition that includes Schmidt Sciences, the UK’s ARIA agency, the Cooperative AI foundation and Google.org. Lead researcher Rohin Shah says the grant aims to create a field for multi‑agent safety before agents flood the market. The initiative follows DeepMind’s showcase of tools at Google I/O, signaling shift toward governance.

Shah warns that as autonomous agents begin to execute tasks without human oversight, their interactions could amplify existing internet threats—scams, prompt‑injection attacks, and coordinated cyber‑intrusions. By dropping agents into sandbox simulations, researchers hope to observe emergent behaviors that single‑agent studies miss. The effort mirrors recent moves by Anthropic and security firms urging zero‑trust approaches to AI deployment.

Cybersecurity veteran Refael Angel of Akeyless applauds the independent funding, arguing that no single lab should dictate safety standards for improvising agents. He cautions researchers not to overlook present‑day abuses while chasing speculative catastrophes. The consortium’s sandbox program therefore represents the first coordinated attempt to map and mitigate real‑world today risks of a burgeoning multi‑agent economy in practice.