'No single safeguard is sufficient': IMDA unveils 3 new initiatives to strengthen AI safety“任何单一保障措施都不足以应对人工智能安全问题”：新加坡资讯通信媒体发展局（IMDA）推出三项新举措，加强人工智能安全防护
Singapore has introduced three initiatives to strengthen artificial intelligence (AI) safety and cybersecurity as increasingly autonomous AI systems pose new risks, including unauthorised access to sensitive information and unintended actions.Announced by the Infocomm Media Development Authority (IMDA) on Wednesday (Oct 14), the initiatives include a testing guide for agentic AI systems, an accreditation scheme for AI testing firms, and a...
Singapore has introduced three initiatives to strengthen artificial intelligence (AI) safety and cybersecurity as increasingly autonomous AI systems pose new risks, including unauthorised access to sensitive information and unintended actions.
Announced by the Infocomm Media Development Authority (IMDA) on Wednesday (Oct 14), the initiatives include a testing guide for agentic AI systems, an accreditation scheme for AI testing firms, and a taskforce to strengthen cybersecurity defences.
Speaking at the AI Assurance Exchange: Agentic AI Edition at Novotel Singapore on Stevens, IMDA chief executive Ng Cher Pong said: "When we give AI more agency, we need greater assurance over how that agency is exercised."
New guide to help organisations test AI agents
The first initiative is the Starter Kit for Testing Agentic Systems for Safety and Reliability, which provides guidance on assessing AI agents before deployment.
Unlike traditional AI applications, agentic AI systems can use tools, access sensitive information and execute tasks on behalf of users, potentially causing unintended consequences.
Ng cited examples including an experimental AI agent that accessed non-public information in an Australian government system, and another that exploited a vulnerability in a gym booking system to remove another person from a waitlist.
A third agent handling a Marketplace transaction shared a seller's home address with a buyer and arranged a visit without the seller's knowledge.
The new guide outlines a four-step testing approach: Identifying risks, designing tests, analysing test results and logs, and implementing continuous monitoring and retesting.
It also emphasises examining an agent's actions throughout a task, rather than assessing only its final output.
The guide draws on work by the Global AI Assurance Sandbox and the Singapore AI Safety Institute, as well as consultations with more than 20 industry participants.
5 firms accredited to test AI applications
The second initiative is the AI Tester Accreditation Programme (AI TAP), launched by the AI Verify Foundation to help businesses identify qualified third-party firms to assess AI safety and reliability.
The first five accredited firms are AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan.
Firms undergo technical, operational and financial evaluations, with accreditation valid for one year.
Ng said accreditation does not mean a firm can test every type of AI system. Instead, firms are recognised for specific testing competencies, helping businesses identify suitable providers for their needs.
He also encouraged more experienced AI testing firms to apply as demand for independent assessments grows.
Taskforce to strengthen cybersecurity
The third initiative is the formation of the Frontier AI Resilience and Security Taskforce (Fairs), which will strengthen the infocomm and media sector's defences against AI-enabled cyber threats.
Bringing together more than 20 leaders and experts from telecommunications, media, cybersecurity and AI, the taskforce will focus on improving responses to AI-accelerated threats, strengthening critical systems against vulnerabilities, and sharing intelligence on emerging attack techniques.
It is expected to publish a report containing practical guidelines and best practices in the first half of 2027.
Ng said the three initiatives address different aspects of AI assurance, from testing individual systems and identifying qualified testers to strengthening cybersecurity across the wider sector.
"Ultimately, the objective is to help organisations adopt AI with greater confidence."
asyiqin.nadzri@asiaone.com
新加坡推出了三项举措，以加强人工智能 (AI) 安全和网络安全，因为日益自主的 AI 系统带来了新的风险，包括未经授权访问敏感信息和意外行为。
新加坡资讯通信媒体发展局（IMDA）于周三（10 月 14 日）宣布了这些举措，其中包括人工智能系统测试指南、人工智能测试公司认证计划以及加强网络安全防御的特别工作组。
在新加坡史蒂文斯诺富特酒店举行的 AI 保障交流会：智能体 AI 版上，新加坡资讯通信媒体发展局 (IMDA) 首席执行官吴哲邦表示：“当我们赋予人工智能更多自主权时，我们需要对这种自主权的行使方式有更大的保障。”
帮助组织测试人工智能代理的新指南
第一项举措是“用于测试智能体系统安全性和可靠性的入门工具包”，它为部署前评估人工智能智能体提供了指导。
与传统人工智能应用不同，智能体人工智能系统可以使用工具、访问敏感信息并代表用户执行任务，这可能会造成意想不到的后果。
吴恩达举例说，其中一个实验性人工智能代理访问了澳大利亚政府系统中的非公开信息，另一个则利用健身房预订系统中的漏洞将另一个人从候补名单中移除。
第三位经手 Marketplace 交易的经纪人将卖家的家庭住址透露给了买家，并在卖家不知情的情况下安排了买家上门看房。
新指南概述了四步测试方法：识别风险、设计测试、分析测试结果和日志，以及实施持续监控和重新测试。
它还强调要考察智能体在执行任务过程中的行为，而不是仅仅评估其最终结果。
该指南借鉴了全球人工智能保障沙箱和新加坡人工智能安全研究所的工作成果，并咨询了 20 多家行业参与者的意见。
5家公司获得认证，可以测试人工智能应用程序
第二项举措是由 AI Verify 基金会发起的 AI 测试员认证计划 (AI TAP)，旨在帮助企业找到合格的第三方公司来评估 AI 的安全性和可靠性。
首批获得认证的五家公司分别是 AIDX Tech、Knovel Engineering、普华永道新加坡、Resaro 和 Vulcan。
企业需接受技术、运营和财务评估，认证有效期为一年。
吴先生表示，获得认证并不意味着一家公司可以测试所有类型的AI系统。相反，认证认可的是公司在特定测试方面的能力，这有助于企业找到满足自身需求的合适供应商。
随着对独立评估的需求不断增长，他也鼓励更多经验丰富的AI测试公司申请参与。
成立工作组加强网络安全
第三项举措是成立前沿人工智能韧性和安全工作组（Fairs），该工作组将加强信息通信和媒体行业抵御人工智能网络威胁的能力。
该工作组汇集了来自电信、媒体、网络安全和人工智能领域的 20 多位领导人和专家，将致力于改进对人工智能加速威胁的应对措施，加强关键系统抵御漏洞的能力，并分享有关新兴攻击技术的情报。
预计将于 2027 年上半年发布一份包含实用指南和最佳实践的报告。
吴先生表示，这三项举措针对人工智能保障的不同方面，从测试单个系统和确定合格的测试人员，到加强整个行业的网络安全。
“最终目标是帮助企业更有信心地采用人工智能。”
asyiqin.nadzri@asiaone.com