Singapore’s Answer to AI Doomsday Warnings: Audit the Systems, Govern the Agents

Singapore’s AI industry voices are treating the renewed warnings about catastrophic risk as a reason to strengthen oversight, not to choose between unchecked development and a halt to all AI work. Their emphasis is on independent evaluation, transparency about safety practices and enforceable limits on what deployed systems can access or do. 732

Why the debate intensified in September

Jacob Coxon resigned from Anthropic after working at both Anthropic and OpenAI, accusing the companies of racing towards self-improving superintelligence without acting responsibly. Anthropic researcher Evan Hubinger publicly agreed with the concern and gave a personal estimate of greater than 10% for AI causing human extinction within the next decade. That figure is his judgment, not a measured probability or evidence that extinction is imminent. 34

Anthropic chief executive Dario Amodei called for slower development. His September 12 proposal also included giving third-party evaluators access to verify safety practices and communications and report incidents. That shifts part of the debate from how frightened should we be? to what can outsiders check? 1332

What independent auditing can—and cannot—settle

Testing can identify limitations in systems that exist today. Singapore’s AI Verify toolkit, launched in 2022, supports assessments of whether AI results are fair, robust and explainable. An AI Tester Accreditation Programme is also described as a way to build testing capacity. Neither amounts to proof that every future system will be safe. 1933

Independent scrutiny requires more than a provider’s assurance that its model passed an internal check. Evaluators need enough access to examine stated practices and surface incidents. Organisations deploying AI also need transparency and controls from providers before connecting models to their data and workflows. Access to model weights alone should not be mistaken for a complete account of training decisions, failures or safeguards. 3234

This is not an argument against slowing any development. One Singapore-based expert cited in the debate argues that developers should pause an activity if they cannot adequately contain its system or explain how serious risks are managed. The distinction is between a condition tied to a demonstrable risk and a blanket pause across applications. 32

Govern agents before giving them authority

Autonomous agents make the oversight question immediate: a system that can use tools and act across workflows needs boundaries on its authority. Singapore’s Infocomm Media Development Authority launched its Model AI Governance Framework for Agentic AI in January 2026. Its guidance calls for assessing risks upfront and limiting agents’ autonomy and access to tools and data; the updated framework specifies least-privilege access. It is guidance, not a blanket legal prohibition on agent deployment. 504852

For an organisation, that means identifying who owns an agent, restricting its permissions, keeping actions traceable and reversible where possible, and monitoring its operation. Those controls are more actionable than a general assurance that an agent is safe. 34 Minister Josephine Teo has compared the trust-building role of AI safeguards and standards with safety practices in civil aviation. 17

The central response is accountability under uncertainty: test what can be tested, allow independent scrutiny and withhold powers that cannot be safely governed. None of those steps resolves the long-term AI risk debate, but each gives developers and deployers a concrete obligation now. 3248