Securing the agentic era: Our commitments and proposed Action Blueprint

AI agents are redefining how we interact with digital technology, with tremendous beneficial potential. From general assistants that can perform day-to-day tasks to specialized tools that accelerate cyber defenses and support scientific breakthroughs, they promise gains in everything from productivity to health. However, distrust toward agentic AI remains an obstacle to adoption. Before individuals and organizations are willing to fully commit resources to deploying AI agents, they must have confidence that agents are secure-by-design and operating with appropriate guardrails.

Building on our TRUST Framework for governance of agents, we believe that a safe, reliable, and secure agentic ecosystem is the foundation of widespread adoption, enabling society to fully reap the benefits of AI agents. We also recognize that security in the agentic era demands cross-industry and cross-sector coordination. That’s why we are proposing the Action Blueprint for Agentic Security – a set of policy and operational recommendations designed to coordinate standards for organizations, developers, security leaders, and government partners to act together to ensure a trustworthy agentic AI ecosystem.

The blueprint outlines six types of actions that stakeholders can take to contribute to the overall security of the agentic ecosystem:

• Standards and protocols. Establish standard definitions and integrate security into core protocols for multi-agent interoperability.
• Evaluation and governance. Launch public sector pilot programs and secure sandboxes for agentic use-cases to inform assessment and governance. Enhance multi-layered security and monitoring for model testing and evaluation.
• Collaboration and information sharing. Coordinate information sharing and sharing of agent security and safety best practices, leveraging existing channels and processes for cybersecurity information sharing when appropriate.
• Upgrade infrastructure and code. Prepare digital public infrastructure, such as citizen-facing portals and government digital services, to receive and interact with agents, and secure code development with agents.
• Resilience and readiness. Equip critical sectors and open-source developers with incident response playbooks for novel agentic threats and enhance baseline security for agent development and deployment.
• Enforce secure-by-design. Embed secure-by-design principles into government procurement of agentic systems and deploy defensive agents for automated vulnerability discovery and remediation.
  

To dive deeper into each of these recommendations, read our newly published blueprint: Securing the Agentic Era: The Action Blueprint for Agentic Security.

At Google, we are invested in meeting new challenges and opportunities with AI, and security is no exception. Billions of people trust Google to keep them safe online by default, and this is a track record we have upheld for more than two decades. We are already using agentic AI innovation to secure and defend systems, detect and remediate vulnerabilities, and protect user privacy.

We have focused on agentic security for years. As part of this long-term effort, we have published our approach to secure AI agents, described how we are securing the AI frontier, developed industry best practices through coalitions, donated secure agentic protocols to the ecosystem, invested in and disseminated research advances, shared our lessons learned from testing the robustness of our models, and spearheaded the Secure AI Framework (SAIF). Our approach to technical AGI safety and security further details how we deal with risk from advanced AI agents.

Highlights of our most recent work on agentic security include: launching defensive agentic offerings such as Wiz and CodeMender, announcing a joint $10M technical research fund focused on multi-agent safety, publishing research on agent control, agent vulnerabilities, and agent-assisted vulnerability management, and developing an Agent Security Assessment for developers to identify security risks and potential mitigations for their agents and agent systems.

Google is committed to actively partner and contribute our resources to help drive collective action to secure the digital ecosystem. We have, and will continue to:

• Increase our investment in AI safety and cyber defense research, with a focus on agents 
• Collaborate with industry partners on agent-related incident reporting and threat intelligence information sharing
• Partner with governments and high priority sectors to strengthen defenses against agentic security threats and deploy agents for cyber defense
• Proactively drive and contribute to the development of international agentic AI security frameworks through multi-stakeholder initiatives such as the Coalition for Secure AI (CoSAI) and bodies like ISO/IEC and NIST
• Invest in agentic security resources and tooling for open source developers and maintainers to increase the resilience of the open source ecosystem