AI agents are moving from experimental tools to everyday enterprise workflows.

Reporting live from AWS re:Invent 2025 in Las Vegas for Irish Tech News, I attended a press-only briefing titled Security and the Rise of AI Agents, where senior AWS leaders Amy Herzog, Chief Information Security Officer, Hart Rossman, Vice President in the Office of the CISO, Gee Rittenhouse, Vice President of Security Services and Neha Rungta, Director of Applied Science outlined how the company intends to manage this transition.

AWS is pushing ahead with autonomous agents, but only within a security model built on long-standing principles: identity, governance, compliance and clear oversight.

What is an AI Agent?

An AI agent is a software system that uses artificial intelligence to carry out tasks autonomously in pursuit of a specific goal. Unlike chatbots that only respond to prompts, an agent can reason, plan and take action across different steps of a workflow. It can use tools such as web services or APIs, monitor its progress and adjust its approach as conditions change. Over time, it can improve its performance based on the data and experience it gathers.

This distinction matters, because the rise of agents raises new questions about accountability, access, oversight and safety.

Security First

AWS chief executive Matt Garman shaped much of the week's discussion. Speaking about the reality facing engineering teams, he noted:

"Every customer wants their products to be secure, but you have trade-offs. Where do you spend your time? Do you improve the security of existing features, or do you ship new ones?"

The briefing returned to this point several times. AWS's position is that strong design-stage security reduces the tension between improvement and innovation. Agents are seen as an opportunity to reinforce security, not dilute it.

AWS Security Agent

One of the major announcements at re:Invent was the preview of AWS Security Agent. The tool brings several security checks forward in the development process. It reviews designs, analyses code, gathers richer signals for incident response and performs penetration testing that reflects real system behaviour rather than generic patterns.

AWS Security Agent is one of the new Frontier Agents introduced at re:Invent, a family of autonomous tools designed to handle multi-step tasks across development, security and operations.

Neha Rungta described the significance of this shift. She called the Security Agent "one of these frontier AI agents, a sophisticated class of AI agents that are autonomous and scalable and can work for long periods without human intervention. Security doesn't have to be an afterthought." She added that AWS is expanding its proof-based assurance tools so teams can understand correctness without being specialists in system logic.

The broader point is that verification needs to be continuous, not episodic.

Guardrails for Autonomy

The panel stressed that agents must operate within strict boundaries. Updated policy controls in Amazon Bedrock AgentCore allow organisations to specify what an agent can do, which systems it can reach and how its actions are logged and reviewed.

Hart Rossman remarked that each major technology shift has increased the demands placed on security teams. With agents running for extended periods and across more systems, the real pressure points now are scale and speed. Guardrails are essential.

The Sandbox Approach

A theme repeated throughout the session was the use of sandbox environments. AWS encouraged organisations to test new agents in isolation before considering production use. This allows teams to observe long-running behaviour, confirm access paths, check escalation rules and understand how an agent reacts under different conditions.

The sandbox was presented as a practical way to build confidence gradually rather than relying on assumptions.

Inside the Press Briefing

Questions focused on monitoring autonomy, preventing agents from widening their sco...