Existing guidelines apply to agentic AI risks; government reviewing safeguards: MDDI

[SINGAPORE] Agentic artificial intelligence systems are creating new dimensions of risks, but existing guidelines and regulations apply, Senior Minister of State for Digital Development and Information Tan Kiat How said.

“We are reviewing how to adapt and strengthen these frameworks and systems to account for the increased autonomy of agentic systems, while also monitoring international developments in this emerging field,” he added, in response to supplementary questions from MPs in Parliament on Wednesday (Oct 7).

This follows Minister for Digital Development and Information Josephine Teo’s written answers to oral questions about Singapore’s digital systems’ resilience against agentic AI on Tuesday evening. Teo had called for a “risk-based and multi-layered approach” to enabling innovation while managing risks.

The questions came after an OpenAI agent hacked into an Australian national healthcare database, in the first known case of AI hacking a government network.

“(In Singapore,) no agency has received a report of a cyberattack against its systems involving an unsupervised AI agent so far,” Teo said.

On Wednesday, Tan noted that existing risk-management processes and oversight apply to agentic AI, alongside principles such as maintaining human accountability and putting in place sufficient controls and guardrails. Existing regulation also include incident reporting frameworks.

Long-standing efforts to strengthen data protection, cybersecurity, and resilience also help protect the systems that AI agents interact with, he added.

Government agencies are running trials to stay abreast of the technology and understand responsible implementation, he said, even as he urged private companies to “take appropriate care and responsibility” when deploying agentic AI.

Sectoral regulators already work closely with operators of critical information infrastructure services to assess necessary safeguards, whether voluntary or not, he said.

SEE ALSO

South Korea investigates possible use of AI in hacks on its banks

MAS rolls out AI risk management guidelines for Singapore’s financial sector

Frontier AI companies are also welcome to take greater responsibility in understanding and managing the risks arising from the systems they develop, Tan added.

This could include identifying emerging capabilities and risks early, putting in place safeguards before these systems are widely deployed, and rectifying issues as soon as possible.

But beyond risks arising from frontier AI models, bad actors can also use the technology for malicious purposes, headded.

Meanwhile, as AI tools become more accessible, employees may use them on their own initiative, he said, adding that organisations should encourage employees to use such tools productively while introducing the right safeguards and being mindful of new risks.

“(The onus is) on Singapore and our organisations, especially those running critical information infrastructure to take the necessary safeguards,” he said.

Multi-prong approach

In her earlier response, digital development and information minister Teo also noted that Singapore needs a “risk-based and multi-layered approach” to agentic AI.

First, in deploying AI, limits must be calibrated based on risk, considering the sensitivity of the data and systems involved, the degree of autonomy given to the AI system, and the severity and remediability of any potential harm.

“The greater the potential harm, the stronger the safeguards should be,” she said.

Stronger safeguards could include more rigorous testing for risks, independently verifiable evidence that safeguards are effective, tighter deployment controls or tighter regulatory oversight.

Second, defences against unintended actions and misuse must be strengthened. This means not just telling AI agents what to do, but also deliberately limiting what they can access and do, conducting appropriate test suites, retaining appropriate human oversight, monitoring their behaviour, and containing the consequences if things go wrong.

It also means strengthening cybersecurity and making effective use of AI for cyberdefence.

Third, detect risks and respond early, Teo said, adding that Singapore’s AI safety institute is building technical capabilities to evaluate advanced AI systems, and develop practical methods for testing and assurance.

Fourth, continue to work with international and industry partners, she said. This includes engaging frontier AI developers to obtain information and access to their models where appropriate.

This is not currently a requirement, because access to a model alone does not paint a complete picture of the risks.

“A rigid requirement could even be counter-productive if it leads companies to limit their cooperation or information-sharing,” she said, noting a preference for collaborative relationships.

Meanwhile, Singapore is drawing on a wider range of sources to identify risks of relevance to itself, from developers to other AI safety institutes and international partners.

“By drawing on these different sources, rather than relying on any single form of access or assessment, we can develop a more comprehensive understanding of emerging risks and the safeguards needed to address them.”