DeepMind Leads $10M Multi-Agent AI Safety Research Initiative
Google DeepMind, alongside Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency (ARIA), and Google.org, announced a research funding call of up to $10 million targeting the safety challenges that emerge when multiple independent AI agents interact at scale. The application deadline is August 8, 2026, with awardees expected to be announced in Autumn 2026.
The initiative identifies four priority research domains. Sandboxes and testbeds focuses on creating reproducible evaluation environments including virtual marketplaces, simulated ecosystems, and multi-organization workflows where multi-agent interactions can be studied safely. Agent network science examines how collective capabilities emerge and scale, and how to detect dangerous population-level properties before they manifest in production. Agent infrastructure targets stress-testing security protocols for identity, reputation, and cross-platform interactions. Oversight and control aims to build monitoring methods capable of managing collective harms at scale.
The funding addresses a fundamental gap in current AI safety research: most evaluations focus on single-model behavior, but real-world deployments increasingly involve networks of independently developed agents. Researchers emphasize that as these systems scale, new collective behaviors and capabilities can emerge suddenly, yet the tools to predict, measure, and monitor these transitions do not exist. The initiative seeks to build this instrumentation before widespread multi-agent deployment makes the problem intractable.
Read more — Google DeepMind
Safe & Secure AI Agent Practices
White House Executive Order Establishes Voluntary Frontier Model Framework
The White House issued an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security" in June 2026, establishing a two-track approach combining innovation promotion with security mandates. The order explicitly rejects overly burdensome regulation in favor of voluntary frameworks and collaborative partnerships for frontier AI model development.
On the security track, the order requires federal agencies to act within 30 days: the Department of Defense must upgrade information system defenses, CISA must release directives to protect civilian federal systems and expand AI-enabled defensive tools, and the Treasury Department must establish an AI cybersecurity clearinghouse with industry collaboration. Within 60 days, a classified benchmarking process must be created to assess AI models' cyber capabilities, and a voluntary framework must be designed allowing developers to determine if their models qualify as "covered frontier models."
The voluntary frontier model framework includes a notable provision granting federal agencies access to frontier models for up to 30 days before public release for security evaluation, with intellectual property and confidentiality protections for participating developers. For AI developers, the practical impact centers on the emerging classification of "covered frontier models" and the incentives structure around voluntary participation, which may influence how teams approach model release timelines and security testing.
Read more — The White House
International AI Safety Report 2026: Capabilities Outpacing Defenses
The International AI Safety Report 2026, led by Turing Award winner Yoshua Bengio and authored by over 100 AI experts with backing from more than 30 countries and international organizations, presents a comprehensive assessment of the current state of AI safety. The report finds that while the number of companies publishing Frontier AI Safety Frameworks has more than doubled since the 2025 edition, significant gaps remain between stated commitments and effective implementation.
The report documents that current AI systems can generate code designed to cause harm and discover software vulnerabilities autonomously, citing an AI agent that placed in the top 5 percent of teams at a major cybersecurity competition in 2025. Researchers have refined techniques for training safer models and detecting AI-generated content, but the report concludes that sophisticated attackers can often bypass current defenses, creating an asymmetry where offensive capabilities advance faster than defensive measures.
For developers building AI-powered applications, the report reinforces that safety frameworks must be treated as engineering requirements rather than compliance checkboxes. The finding that defenses are systematically outpaced by capabilities suggests that defense-in-depth approaches combining input validation, output filtering, capability sandboxing, and continuous monitoring remain essential even as individual techniques improve.
Read more — International AI Safety Report
