Outcome Statement – Tabletop Exercise on Agentic AI Security and Control at the Singapore International Cyber Week 2026

• Senior cyber policymakers, critical infrastructure leaders, and AI safety and security experts convened at SICW 2026 to examine the operational security of autonomous AI agents through a scenario-based exercise.

• Participants highlighted the need for system-level safeguards to address the risks of increasingly autonomous agents, while identifying potential areas for public-private and international collaboration.

SINGAPORE, 14 October 2026 – At the 11th Singapore International Cyber Week (SICW 2026), the Cyber Security Agency of Singapore (CSA) and Google convened senior cyber policymakers, critical infrastructure leaders, and AI safety experts for scenario-based exercise on agent security. The exercise aimed to build a shared understanding of developments in agentic AI and their security implications, discuss best practices and policy priorities, and strengthen public-private partnerships on agentic AI security.

Participants discussed the need for greater international alignment between public and private stakeholders. This includes the value of AI to help cyber defenders to detect, analyse, and patch vulnerabilities at machine speed. The ultimate goal of addressing security risks in AI systems is to build trust and confidence so that society can harness the full potential of AI.

Participants examined how the rapid expansion of autonomous AI agents and their growing cyber capabilities is reshaping the operational landscape. Organisations face new challenges as they entrust core functions to autonomous AI agents that operate at machine scale. Relying solely on perimeter defenses or manual human review can lead to inadequate coverage or operational bottlenecks. This makes it more difficult to detect unintended agent behaviour, compromised AI agents, or unexpected multi-agent coordination. Where AI agents perform highly privileged tasks such as monitoring and cybersecurity, organisations should also consider the risks of over-reliance on a single model. Independent verification and model diversity can strengthen resilience.

To address these risks, participants discussed practical safeguards, inspired by industry research on agent control. Rather than assuming models will always behave as intended, safeguards should include system-level (“outside the model”) protections – including continuous runtime monitoring, hardened execution environments, and strict access controls. For deployers, effective oversight requires a layered approach. For example, real-time, deterministic guardrails can prevent unauthorised high-impact actions, combined with asynchronous monitoring to audit agent reasoning and tool use at scale.

The exercise also surfaced practical recommendations, such as the need to prepare for the rapid escalation of novel agentic threats through short-term incident response plans and long-term resilience measures. Other recommendations included standardising protocols around verifiable agent identities and transitioning to contextual, machine-speed access controls, such as those outlined in Google’s “Beyond Zero ↗ (opens in new tab)” security framework.

Participants raised three areas for further discussion and collaborative exploration:

1. Collaborative information sharing: Establishing trusted channels for stakeholders to share threat intelligence and report observations of emerging operational anomalies and unexpected agent behaviours.

2. AI security and testing: Using AI stress-testing, behaviour evaluations, security guardrails, and pilot testing models to ensure the safety and security of autonomous agents and multi-agent interactions before widespread deployment.

3. Incident response and human oversight: Investing in preparedness and capacity for incident response while navigating resource limitations, legal and compliance obligations in decision-making, and ensuring appropriate human oversight of consequential actions and decisions.

Participants welcomed the constructive exchange between cybersecurity policymakers, industry leaders, and AI safety experts. They reaffirmed the importance of sustained international co-operation to build a secure, resilient, and trustworthy agentic ecosystem.