Google DeepMind, in partnership with Schmidt Sciences, the UK’s ARIA, the Cooperative AI Foundation, and Google.org, has established a ten million dollar research initiative to investigate the systemic risks of large-scale artificial intelligence agent interactions. The program aims to cultivate independent academic and institutional research into multi-agent safety, a discipline that currently lacks standardized frameworks. Rohin Shah, head of AGI safety and alignment at Google DeepMind, and James Fox of Schmidt Sciences emphasize that as autonomous agents proliferate and begin operating collaboratively without constant human oversight, the digital infrastructure faces unprecedented vulnerabilities. The consortium’s primary concern centers on the compounding effects of existing cyber threats when scaled across millions of interacting agents. Risks include automated fraud, prompt injection attacks that repurpose agents into malicious runtime software, and coordinated cyber operations. Shah notes that while a complete economic collapse is unlikely in the immediate term, the industry is racing against a potential deployment tipping point estimated to occur within months. To mitigate these threats, researchers will prioritize realistic sandbox simulations, moving beyond isolated testing to observe emergent behaviors in mass-interaction environments. Fox adds that predicting outcomes requires modeling the non-deterministic dynamics inherent in large-scale agent networks, where collective capabilities may eventually surpass individual systems. The initiative aligns with broader industry shifts toward proactive security protocols. Anthropic recently introduced a zero-trust deployment framework for AI agents, assuming inherent system vulnerabilities and potential adversarial behavior. Refael Angel, co-founder and CTO of Akeyless, welcomes the collaborative funding model, arguing that security standards must be collectively developed rather than dictated by single laboratories. He cautions against chasing theoretical threats while neglecting immediate, practical vulnerabilities, noting that traditional cybersecurity models built on static code execution are fundamentally incompatible with AI systems capable of real-time reasoning and adaptive execution. By funding independent research and establishing simulation-based evaluation protocols, the consortium aims to transform multi-agent safety from a conceptual concern into an operational discipline. The project underscores a growing consensus that the safety of interconnected AI ecosystems requires preemptive, scalable strategies before autonomous systems reshape digital and economic landscapes.