Google Cloud flags AI agents' security & governance risks

Google Cloud has published a report on AI infrastructure that identifies security, governance and operations as the main barriers to scaling AI inference, with a particular focus on the rise of AI agents in business systems.

It finds that 79% of technology leaders see security, governance or operations as their biggest challenge in expanding inference workloads. It also highlights growing concern over how organisations control AI agents that can read emails, query databases and trigger application programming interface calls.

Unlike earlier software tools, AI agents do more than retrieve information. They can take actions across connected systems, increasing the risks around permissions, identity management and oversight.

One of the clearest deployment obstacles is multi-system access. The report says 35% of senior IT decision-makers cite inadequate security for access across multiple systems as a primary issue holding back agentic deployment.

The research describes this as a tension between usefulness and control. Agents need broad access to complete tasks, but each new connection expands the area security teams must protect.

New risk model

That shift is changing the threat model for businesses adopting autonomous workflows. The report points to risks including tool poisoning and indirect prompt injection, where an attacker manipulates an agent's behaviour through the data it processes.

It argues that legacy security approaches are poorly suited to these automated threats. Managing dynamic permissions for agents is becoming a growing challenge for companies whose access controls were designed for human users and more static software environments.

Security concerns extend beyond identity. Organisations must secure both the network layer and the model layer, while security teams are placing greater emphasis on verifying provenance and guarding against misuse, rather than focusing only on breach prevention.

Platform choice

The findings suggest many companies are responding by seeking more centralised control. Google Cloud says 69% of surveyed executives now view a full-stack platform as a critical requirement, while 80% say data compliance is the main factor shaping that choice.

That preference reflects a broader push for integrated oversight of AI systems, data access and operational controls. For businesses deploying agents into sensitive processes, governance structures are becoming a prerequisite rather than a later-stage consideration.

Google Cloud links that approach to frameworks such as its Secure AI Framework, or SAIF, and to the use of a central control plane to manage agent activity. Organisations are increasingly looking for tools that can apply policy, identity rules and monitoring in one place.

The report highlights three areas of risk management. The first is secure-by-default design, which builds security controls into the AI development process from the outset and aims to reduce exposure to threats such as prompt injection.

The second is governance and oversight tailored to agents, including permission structures and identity controls designed to limit risky interactions and improve visibility into what an agent can access and do.

The third is human review for critical actions. In practice, that means setting rules that require human approval before an agent can take high-impact steps.

Governance focus

The report presents governance not simply as a compliance burden, but as a condition for wider use of AI in core business functions. The argument is aimed at organisations trying to extend AI from pilot projects into operational systems, where the consequences of error or misuse are greater.

For cloud providers and enterprise software groups, the issue has commercial significance as customers assess whether existing security tools can cope with systems that act with greater autonomy. The emphasis on integrated platforms also reflects a market trend in which buyers want fewer gaps between model management, infrastructure controls and security operations.

By framing AI agents as "ultimate insiders", the report underlines the degree of trust businesses may place in these systems. It argues that this trust must be matched by stricter governance over access, behaviour and escalation paths when an agent moves beyond information retrieval into decision-making and execution.

The document concludes that organisations deploying agents across sensitive workloads will need to rebuild parts of their technology stack around security and governance if they want to expand AI use without increasing operational risk.