Securing Agentic AI: An Addendum to the Guidelines and Companion Guide on Securing AI Systems 2025 (Draft for Public Consultation)
SECURING AGENTIC AI An Addendum to the Guidelines and Companion Guide on Securing A I Systems 2025 (DRAFT FOR PUBLIC CONSULTATION)
This document provides practical mitigation measures and security controls for agentic artificial intelligence systems. Published as a draft addendum for public consultation by the Cyber Security Agency of Singapore, it builds upon foundational AI security principles by addressing risks across the development lifecycle, system workflows, and agent capabilities.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
System owners, decision-makers, AI practitioners, and cybersecurity professionals should read this to secure agentic AI deployments.
What the document contains
The document addresses cybersecurity threats, baseline components, architecture design, and capabilities of agentic AI systems. It outlines practical treatment measures and security controls across planning, development, deployment, operations, and maintenance.
The guide does not cover AI safety topics such as fairness, transparency, or inclusion, nor does it address the misuse of AI for cyberattacks or scams like deepfakes.
- Agentic AI systems operate with reduced human oversight and introduce novel security considerations around system boundaries and rogue actions.
- System owners should use a lifecycle approach combined with risk assessments and taint tracing across agentic workflows.
- The document categorizes agentic risks across baseline components, system designs, and cognitive, interaction, and operational capabilities.
- Practical case studies illustrate the deployment of safeguards in SaaS applications, client onboarding systems, and automated fraud detection.
- Protocols such as the Model Context Protocol and Agent2Agent framework present specific interoperability and security considerations.
· Quick Reference Table · 1. Introduction · 2. How Agentic AI Works · 3. Security Threats to Agentic AI Systems · 4. Securing Agentic AI · 5. Use Case Example · Annex A Threats to Agentic AI Systems · Annex B Model Context Protocol · Annex C Agent 2 Agent Protocol · References
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Accountability
2/5Clear responsibility for AI outcomes
Roles and access controls for agentic AI systems should be clearly defined to avoid unauthorised access or excessive privilege.
2.2.2. Roles & access control
screening score · 2026-10-23
Fairness
1/5Non-discrimination and equitable treatment
It does not specifically address AI safety, or other common attendant considerations for AI such as fairness, transparency or inclusion
10. PURPOSE AND SCOPE
screening score · 2026-10-23
Human Dignity
0/5Respect for the inherent worth of persons
screening score · 2026-10-23
Human Oversight
4/5Humans remain in control of consequential decisions
Ensure adequate human oversight (human-in-the-loop) to verify model or agent output, when viable or appropriate.
4.4. ENSURE ADEQUATE HUMAN OVERSIGHT (HUMAN-IN-THE-LOOP)
screening score · 2026-10-23
Privacy
2/5Protection of personal data and private life
Agents may inadvertently store sensitive user or organisational data from prior interactions, resulting in data privacy risks.
2.4 Identify, Track and Protect AI system assets
screening score · 2026-10-23
Transparency
1/5Openness about how AI systems work and are used
It does not specifically address AI safety, or other common attendant considerations for AI such as fairness, transparency or inclusion
10. PURPOSE AND SCOPE
screening score · 2026-10-23
Enforcement
1/5Consequences and compliance mechanisms with real force
These measures and controls are voluntary, and not all the measures and controls listed in this addendum will be applicable to all organisations
10. PURPOSE AND SCOPE
screening score · 2026-10-23
Impact Assessment
4/5Required assessment of an AI system's risks/impacts before or during use
Conduct a risk assessment, either based on best practices or your organisation’s existing Enterprise Risk Assessment/Management Framework.
4.1. TAKE A LIFECYCLE APPROACH, AND START WITH A RISK ASSESSMENT
screening score · 2026-10-23