No report of rogue AI agents attacking S’pore government agencies: Josephine Teo

SINGAPORE - Local government agencies have not received reports of attacks on their systems by unsupervised artificial intelligence agents, said Minister for Digital Development and Information Josephine Teo in a written parliamentary reply on Oct 6.

Even so, firms should not passively wait to be informed about risks and breaches, Senior Minister of State at the Ministry of Digital Development and Information Tan Kiat How told Parliament on Oct 7.

“We cannot just wait for other people to tell us that they are hacking us. The onus is on Singapore and our organisations, especially those running critical information infrastructure, to take the necessary safeguards,” Tan said.

WP MP Fadli Fawzi had asked if public agencies have robust safeguards against potential breaches by AI agents, and if any of such breaches had taken place. MP Alex Yam’s (Marsiling-Yew Tee GRC) had also asked whether frontier AI developers are obligated to notify the authorities about cyber incidents promptly.

Such attacks by rogue agents from frontier AI labs have made headlines in recent months, following successful attempts by AI agents to escape secured environments.

“We will continue to monitor developments, including incidents overseas, and apply relevant lessons to strengthen our safeguards and reporting arrangements,” said Teo in her written response.

Noting that Singapore has joined the international call for stronger safeguards on frontier AI, Teo said that the local authorities also engage frontier AI developers to obtain information and access to AI models where appropriate.

“We currently do not make this a requirement, because access to a model by itself does not necessarily give 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,” said Teo.

Reiterating her point, Tan said on Oct 7 that cyber threats can also stem from bad actors using open-weight AI models. Unlike proprietary models from US firms such as OpenAI and Anthropic, open weight models are publicly available for anyone to run and modify.

“There are many actors out there who can download open-weight models, adapt them, create their own harness and context, and use them for bad purposes,” said Tan.

While acknowledging the risks, firms here should not be paralysed by the doomsday narratives. With the right safeguards and cyber hygiene, they could benefit from AI by using the tech to create new products and services, he said.

Alarms over rogue AI’s hacking abilities were sounded when ChatGPT maker OpenAI disclosed in July that one of its AI agents had earlier escaped its testing environment and breached AI software repository Hugging Face to complete a task it was given. The attack was not intended by OpenAI.

In September, Australia said that OpenAI’s AI agent had breached its government health data portal in June, and gained unauthorised access to public and non-public files such as statistics on public medical spending.

Concerns over the speed of development of AI have sparked intense debate among tech leaders, safety researchers, and politicians on the need for regulation. While US president Donald Trump has labelled renewed safety concerns about advanced AI models as a “hoax”, he also oversaw the signing of a voluntary safety pact among US tech executives that pledged stronger safeguards over AI systems.

Singapore will neither dismiss the potential “catastrophic” or “extinction-level” risks that AI poses, nor assume that every scenario will materialise, said Teo in her written reply on Oct 6.

“Our approach is therefore to monitor the evidence closely, build the technical capabilities needed to understand advanced AI systems, and work internationally on measures to address severe risks as the evidence develops,” said Teo.

Within the public service, the use of AI agents will take into account the sensitivity of data and systems involved, actions AI agents are allowed to take, the severity of potential harm and if the harms can be remedied, said Teo.

For the wider economy, Singapore has developed resources to provide guidance on managing risks and facilitate real-world testing of safeguards. These include the Infocomm Media Development Authority’s Model AI Governance Framework for Agentic AI, and the Global AI Assurance Sandbox.

The Singapore AI Safety Institute is also working on building technical capabilities to evaluate advanced AI systems, said Teo.

She was responding to MP Valerie Lee’s (Pasir Ris-Changi GRC) question on the institute’s testing skills, and MP Charlene Chen’s (Tampines GRC) query on whether existing incident-reporting requirements adequately capture serious incidents involving AI systems.

“After AI systems are deployed, we must monitor their behaviours and learn from incidents and near-misses when they occur,” said Teo.

She said that cyber incidents, including those involving AI, should continue to be reported to the Cyber Security Agency of Singapore’s Singapore Cyber Emergency Response Team, and GovTech’s Vulnerability Disclosure Programme.

“We are looking into how these existing channels can be better leveraged to identify incidents involving AI, support remediation and improve safeguards, and whether further coordination or reporting arrangements are needed.”

Existing incident reporting thresholds remain relevant for AI-related incidents, said Tan. These include mandatory breach notification for data leaks involving 500 or more individuals under the Personal Data Protection Act.

Referencing a recent AI-related data breach where more than 95,000 Bee Cheng Hiang customers had their e-mail addresses exposed after an employee used a bad prompt in an AI tool, Tan said that this incident highlighted the need for organisations to manage the risks associated with “shadow AI”.

This refers to the use of AI tools by employees which are not approved in the workplace.

“We should not discourage such productive use of AI, but organisations need to be aware of how these tools are being used and put appropriate policies and safeguards in place.”