Cybersecurity policymakers, critical infrastructure leaders and AI security experts have called for stronger safeguards, independent verification and international cooperation to address the growing operational risks associated with autonomous artificial intelligence.
The Cyber Security Agency of Singapore (CSA) and Google convened senior cybersecurity policymakers, critical infrastructure representatives and AI safety experts to examine the security implications of increasingly autonomous artificial intelligence systems.
Held during the 11th Singapore International Cyber Week (SICW 2026), the tabletop exercise explored how autonomous AI agents could affect cybersecurity operations, critical infrastructure protection and organisational risk management.
According to an outcome statement released on 14 October 2026, participants examined scenarios involving AI agents capable of independently performing complex tasks, accessing sensitive systems and coordinating with other agents at machine speed.
The exercise highlighted concerns that traditional cybersecurity controls, including perimeter defences and manual human oversight, may be insufficient as organisations increasingly delegate operational responsibilities to autonomous AI systems.
Participants identified several emerging risks, including unintended agent behaviour, compromised AI agents, unexpected interactions between multiple autonomous systems and excessive reliance on individual AI models.
The discussions also examined the implications of deploying AI agents with privileged access to enterprise systems, particularly where they perform cybersecurity monitoring, vulnerability management or other sensitive operational functions.
A central finding was the need to introduce security controls operating independently of the AI models themselves.
Rather than assuming autonomous systems will consistently behave as intended, participants recommended implementing system-level safeguards, including continuous runtime monitoring, hardened execution environments and strict access controls.
The exercise considered a layered security approach combining real-time, deterministic controls capable of preventing unauthorised high-impact actions with additional monitoring to examine agent reasoning, decision-making and the use of external tools.
Such arrangements could help organisations identify potentially harmful behaviour without relying entirely on the AI system to assess its own actions.
Participants also discussed the importance of independent verification and model diversity, particularly in environments where organisations depend on a single AI model to perform highly privileged functions.
The outcome statement warned that excessive dependence on one model could introduce additional operational vulnerabilities, making independent verification an important consideration for system resilience.
Another priority was the development of stronger identity and access management mechanisms for autonomous agents.
Participants discussed standardising protocols for verifiable agent identities and introducing contextual access controls capable of operating at machine speed.
The discussions referenced Google’s Beyond Zero security framework as an example of an approach intended to support more dynamic security controls for autonomous systems.
The exercise also acknowledged the potential benefits of AI-enabled cybersecurity, including the ability to detect, analyse and remediate vulnerabilities at speeds beyond conventional human-led operations.
However, participants emphasised that these capabilities would require corresponding improvements in security governance, monitoring and incident response.
Three areas were identified for further international and public-private collaboration.
The first involved establishing trusted information-sharing arrangements to exchange intelligence about emerging threats, operational anomalies and unexpected AI agent behaviour.
The second focused on AI security testing, including stress testing, behavioural evaluations, security guardrails and controlled pilot deployments before autonomous systems are introduced into production environments.
The third addressed incident response and human oversight, particularly the need to prepare for rapidly escalating incidents involving autonomous agents while maintaining appropriate accountability for consequential decisions.
Participants acknowledged that organisations would need to balance operational efficiency with legal, regulatory and compliance obligations, particularly when AI agents are authorised to take actions affecting critical systems or sensitive information.
The exercise also highlighted the importance of preparing both immediate incident response arrangements and longer-term resilience strategies to address emerging threats.
For Australian organisations, the findings are relevant to the growing adoption of agentic AI across government, financial services, critical infrastructure and enterprise cybersecurity operations.
As AI systems move beyond advisory functions into autonomous execution, security teams face new challenges in defining permissions, validating machine-generated decisions and ensuring that automated actions remain within authorised operational boundaries.
The discussions suggest that conventional approaches to identity management, privileged access and security monitoring may require adaptation to accommodate autonomous agents capable of executing tasks and interacting with multiple systems without continuous human intervention.
The CSA and Google initiative reflects broader international efforts to establish practical security expectations for agentic AI before the technology becomes deeply embedded in critical business and government operations.
Participants concluded that sustained cooperation between governments, technology providers, cybersecurity practitioners and AI safety researchers would be essential to developing a secure and resilient environment for autonomous AI.
The outcome statement did not announce new mandatory regulatory requirements or a formal international agreement. Instead, it identified operational safeguards and areas for further collaborative work as organisations prepare for wider deployment of increasingly capable AI agents.
