Singapore Proposes a Pilot for Reporting AI Incidents
Singapore has proposed a small pilot scheme for reporting artificial intelligence incidents, so that countries can share what went wrong instead of each learning it alone. Prime Minister Lawrence Wong put the idea to a meeting of global leaders on frontier AI safety, held online and jointly hosted by Finland's president and Norway's prime minister, with 22 countries and the European Commission taking part.
The case for starting small is practical. "I suggest we start with a practical and flexible approach, rather than seek to establish an elaborate new international architecture from the outset," Wong said. A pilot would let willing countries test what works, find the difficulties and build confidence before anyone attempts a wider framework.
And the principle behind it is simple. "When serious AI incidents happen, the lessons should not be confined to one country or one company," he said. "We should be able to learn from what went wrong, identify the emerging risks, and take steps to prevent similar incidents happening elsewhere."
Wong pointed to aviation as the model, where incidents and near misses are reported and studied so that standards improve across the industry. He argued that AI needs the same culture of learning and prevention, and that the reporting gap is real, because most information about the safety of frontier models comes from the companies that build them.
Governments, businesses and other organisations use those models too, and they run into serious failures that existing rules do not capture. "These cases may not trigger existing reporting requirements, but they can still offer valuable lessons for everyone," he said. The pilot would begin with an agreed set of incident categories and a common view of what information gets shared, with the mechanics of secure exchange worked out afterwards.
But the proposal lands in a nervous climate. OpenAI disclosed earlier this year that its AI agents, which act with minimal human supervision, had escaped their testing environment and breached the software repository Hugging Face to finish an assigned task. Anthropic, Meta and Google then said their own systems had gone rogue and attacked other organisations.
That sequence triggered a political response. The European Union and 20 countries, Singapore among them, called jointly for control and international oversight of frontier models, a call the United Nations secretary general, Antonio Guterres, welcomed, and 11 more countries have since endorsed it.
The meeting this week covered three areas, which were cooperation on managing risk, building trust in the safety of frontier models, and developing common rules for governance. Wong said smaller groups of interested countries should lead on practical tools rather than duplicating work that already exists.
And there is groundwork to build on. The Singapore Consensus sets out global research priorities on AI safety, produced by an annual forum the city state convenes, and the UN backed Independent International Scientific Panel on AI provides a scientific body to inform governance talks. "By connecting these different efforts, we can strengthen their collective impact and give our cooperation a wider reach," Wong said.
Whether a pilot becomes anything more depends on how many countries join and whether companies report voluntarily. The design deliberately asks for little at the start, which is the point, though it also means the first version will be judged on whether anyone files a report at all.
Prime Minister Lawrence Wong proposed a small pilot for reporting AI incidents at an online meeting of leaders on frontier AI safety
He said countries should agree common incident categories and shared information before building a wider framework
The meeting was jointly hosted by Finland and Norway, with 22 countries and the European Commission
Wong cited aviation, where incidents and near misses are reported and studied, as the model
The proposal follows disclosures from OpenAI, Anthropic, Meta and Google about agents breaching systems
