Guidelines and Companion Guide on Securing AI Systems

Summary

Voluntary cybersecurity guidelines of Singapore's Cyber Security Agency for owners of AI systems, which should be secure by design and secure by default. They identify threats such as supply-chain attacks and adversarial machine learning, call for a security risk assessment and set out measures across five lifecycle stages: planning and design, development, deployment (incident management, security checks before release), operations and maintenance (monitoring, vulnerability disclosure) and end of life. A Companion Guide lists practical controls, and a 2026 addendum covers agentic AI. AI safety, fairness and transparency are outside their scope.

Safety topics

Status notes

CSA consulted the public from 31 Jul to 15 Sep 2024 and received 28 submissions, and launched the Guidelines and Companion Guide (version 1.0 of 15 Oct 2024) at Singapore International Cyber Week on 15 Oct 2024. The Guidelines draw on the UK NCSC and US CISA Guidelines for secure AI system development, the NIST Risk Management Framework, MITRE ATLAS and the OWASP Machine Learning Security Top 10; they also leave out the misuse of AI in cyberattacks, disinformation and scams. Neither document is mandatory; CSA maintains both as living resources. The addendum 'Securing Agentic AI', released for consultation on 22 Oct 2025 (to 31 Dec 2025) and published as version 1.0 on 17 Jun 2026, is to be read with them: it maps agentic workflows to find where threat actors could exploit vulnerabilities, lists agentic AI risks and controls by capability across the lifecycle, and gives examples such as coding assistants, automated client onboarding and fraud detection.

Sources

Open the official source https://www.csa.gov.sg/news-events/press-releases/launch-of-guidelines-and-companion-guide-on-securing-artificial-intelligence-systems/

Secondary source: https://www.csa.gov.sg/resources/publications/guidelines-and-companion-guide-on-securing-ai-systems/

Local copy

Verification

History

1. 2026-10-18notes: Public consultation ran 31 Jul-15 Sep 2024 (28 submissions). Launched at Singapo → CSA consulted the public from 31 Jul to 15 Sep 2024 and received 28 submissions, — Companion Guide version, scope exclusions, the not-mandatory character and the 2026 addendum added; version 1's remark that later addenda were not verified is about the research, not the instrument, and was removed2026-10-18-recheck-south-asia-1 · Claude Code (claude-opus-5-5)
2. 2026-10-18safety_topics: ["evaluation & testing"] → ["evaluation & testing", "incident reporting"] — version 1 had one topic, below the minimum of two; incident reporting added: the Guidelines require incident management procedures at deployment (section 3.2) and a vulnerability disclosure process (4.4)2026-10-18-recheck-south-asia-1 · Claude Code (claude-opus-5-5)
3. 2026-10-18summary: Voluntary cybersecurity guidelines for AI. They identify AI-specific threats suc → Voluntary cybersecurity guidelines of Singapore's Cyber Security Agency for owne — rewritten from the Guidelines' own text (stored in this run): secure by design and default, risk assessment, lifecycle measures, not mandatory, scope excluding AI safety, fairness and transparency; the addendum mentioned2026-10-18-recheck-south-asia-1 · Claude Code (claude-opus-5-5)
4. 2026-10-18date_last_updated: – → 2026-06-17 — CSA's addendum 'Securing Agentic AI – An Addendum to the Guidelines and Companion Guide on Securing AI Systems' was officially published as version 1.0 on 17 Jun 2026 (stored in this run); it extends this instrument and is to be read with it. The Guidelines and Companion Guide texts themselves are unchanged since October 2024.2026-10-18-recheck-south-asia-1 · Claude Code (claude-opus-5-5)

Related records

Singapore MGF for Generative AI Singapore · Published (non-binding, current)

Citation

Cyber Security Agency of Singapore (CSA). Guidelines and Companion Guide on Securing AI Systems. 2024-10-15. https://www.csa.gov.sg/news-events/press-releases/launch-of-guidelines-and-companion-guide-on-securing-artificial-intelligence-systems/. Status: Published (non-binding, current); verified 2026-10-18 (level A). Local copy archive/SG-06/2026-10-18/official.html retrieved 2026-10-18 00:48:34 UTC, SHA-256 c6795cf44a3a16dc3bfd6cf174b4bcd543b5db43b9ce13f5ea8711ce629737a5. Every AI Policy Ever, version 1.18 (snapshot 2026-10-18), record SG-06.

Improve this record

Know of an amendment, a status change or an error? Report news or an error on GitHub; the form opens with this record filled in.