AI Incident Reporting, learning from the near miss
In nine days, Singapore took up the near miss, the AI failure that harms no one and so goes unrecorded, at three levels: an international pilot for sharing incidents, a study of how domestic reporting should adapt, and a rule for when government AI agents need a human to approve. Read here from the primary sources: what each step sets up, what comes next, and why the near miss is the hard case.
Most reporting rules start from harm. A data breach, an outage or an attack crosses a threshold, and someone has to tell a regulator. A near miss is the event that crosses no threshold: the agent that nearly sent the wrong message, the payment stopped one step before it went out. Nothing was harmed, so nothing is reported, and the lesson stays with whoever happened to see it.
Between 28 October and 5 November 2026, Singapore took that problem up three times, at three different levels:
- 28 Oct · International The Prime Minister proposes a pilot for reporting AI incidents across countries Remarks at the Virtual Leaders' Meeting on Frontier AI Safety, co-hosted by Norway and Finland ↗ [01]
- 4 Nov · Domestic The ministry names the gap in Parliament and asks IMDA and CSA to study it MDDI written reply to MPs Melvin Yong and Sharael Taha ↗ [02]
- 5 Nov · Government agents GovTech sets out which actions by government AI agents need human approval Opening address by SMS Tan Kiat How, GovTech STACK 2026 ↗ [03]
This is a reading note on those documents, not an evaluation. None of them reports how often AI near misses happen or whether reporting them reduces harm. That data is what a reporting system would produce, and the note does not pretend to have it. What the documents do show is a government working on the same gap from three directions at once.
Rules that wait for harm
Two MPs asked the same question from different sides. Melvin Yong asked whether the ministry would consider mandatory reporting for serious incidents and near misses involving autonomous AI. Sharael Taha asked whether existing cybersecurity frameworks capture AI-agent near misses at all, and how that intelligence could be shared across agencies and critical-infrastructure operators [02].
The written reply starts with what already exists: Singapore has "established reporting requirements for serious incidents involving personal data, cybersecurity and disruptions to critical or regulated services". Then it names the gap plainly. Autonomous AI incidents "may fall outside existing reporting thresholds, particularly when they do not result in reportable harm" [02].
The gap is not particular to Singapore. It is built into how incident regimes usually work. The EU AI Act, for example, defines a serious incident as one that "directly or indirectly leads to" one of four kinds of harm: death or serious harm to health, serious and irreversible disruption of critical infrastructure, an infringement of fundamental-rights law, or serious harm to property or the environment [06]. An event that leads to none of them is, by definition, not a serious incident.
On next steps, the reply sets out a measured approach. "IMDA and CSA are therefore studying how to strengthen the reporting of serious AI-related incidents, including whether additional reporting or information-sharing arrangements are needed." Anything new "should be targeted and proportionate", and must account for "legitimate security and commercial sensitivities". It would build on existing channels, such as CSA's intelligence exchanges with major technology providers and critical-infrastructure operators [02].
So rather than legislate first, the government is studying first and building on channels that already work. The reply gives no timeline. The order is the point: learn what AI incidents look like before deciding what must be reported.
Users as the ones who see it
A week earlier, the Prime Minister had taken the same argument to an international audience, and added an idea of his own: that users, not only developers, are the ones who see these failures [01].
"Today, much of what we know about the safety of frontier models comes from the developers themselves. But in fact, all of us – governments, businesses and other organisations – are using these models. And as users, we ourselves encounter serious failures or near misses that reveal significant risks, even if no actual harm has occurred."
Those cases, he said, "may not trigger existing reporting requirements, but they can still offer valuable lessons for everyone". The model he pointed to was "international aviation, where incidents and near misses are reported and studied, and lessons are shared" [01].
The proposal is practical and staged:
- 1Interested countries "come together to launch a pilot around AI incident reporting".
- 2They "start by agreeing on a common set of incident categories and information to be shared".
- 3They "explore how participating countries can collect and exchange this information securely".
- 4The pilot would "test what works, identify the practical difficulties, and build confidence before considering a wider framework".
Participation would be voluntary, organised as specific workstreams "with smaller cross-regional groups of interested countries", and could connect to the Independent Scientific Panel on AI that supports the UN Global Dialogue on AI Governance. As the country proposing it, Singapore said it "stands ready to do our part" [01].
The order of the steps is the strongest part of the design. Aviation reporting works because everyone agrees on what counts as an incident, and AI has no such agreement yet. So the first step is agreeing on the categories, before anything is collected.
Drafting versus sending
The third document is about prevention rather than reporting, but it draws the line the other two leave open: which AI actions are risky enough to matter [03].
Senior Minister of State Tan Kiat How put the distinction in one sentence: "An assistant that drafts a wrong message is different from an agent that sends it." GovTech's Agentic Risk and Capability (ARC) Framework "connects an agent's capabilities to appropriate safeguards". For "higher-risk actions, including financial transactions, official communications and data deletion", agents need "stronger controls such as human approval". A Standard built on this will guide Desk, GovTech's assistant for public officers, in deciding which actions proceed on their own and which wait for a person [03].
The speech frames the aim as "neither maximum autonomy nor maximum control". It applies to government agents, and the word "including" means the three categories are examples, not a closed list.
There is a further link worth drawing, though the documents do not draw it themselves. A human approval step is also the place where near misses become visible. When an officer declines an action an agent proposed, that is a recorded case of an agent about to do something a person judged wrong. If those refusals were logged and shared, the approval gates of 5 November could become the first supply of near-miss data for the reporting work discussed on 4 November, and an early contribution to the pilot proposed on 28 October.
What is in motion, and what comes next
Read side by side, the three documents describe a staged approach: propose abroad, study at home, and set firm rules first where the risk is clearest, in the government's own agents.
| Item | In motion | Next to settle |
|---|---|---|
| International pilot28 Oct · PMO [01] | A staged, voluntary pilot proposed by Singapore, which "stands ready to do our part". | Which countries join, how the pilot is resourced, and when it starts. |
| Domestic reporting4 Nov · MDDI [02] | IMDA and CSA studying whether more reporting or sharing is needed, building on existing CSA channels. Any change to be "targeted and proportionate". | Whether reporting becomes mandatory, how a near miss is defined, and the timeline. |
| Government AI agents5 Nov · GovTech [03] | Human approval for higher-risk actions, through the ARC framework and a Standard guiding Desk. | Which actions beyond the named examples need approval, whether the approach extends to the private sector, and whether refused actions are logged. |
The same sitting produced two related written replies. On agentic AI assurance, the ministry cited the AI Verify Foundation's AI Tester Accreditation Programme and Singapore's proposal for ISO/IEC 42119-8 on testing generative AI [04]. On AI agents copying themselves across devices, it said "recent research" shows replication under preset conditions, not easy real-world spread [05].
Who else is moving
Malaysia is the nearest comparison. Reporting the Digital Minister's answer in the Dewan Rakyat on 29 October, Bernama said the AI Governance Bill "has been completed and will be submitted to the Cabinet in the near future", with tabling planned for early 2027. The minister said responsibility "should be determined based on one's role, level of control and the degree of risk", and the report says the bill covers procedures for reporting and handling incidents [07]. The bill's text is not public, and this account comes from the state news agency's report rather than the parliamentary record, so its incident provisions cannot yet be read.
ASEAN concluded negotiations on its Digital Economy Framework Agreement in May 2026, with signing planned alongside the 49th ASEAN Summit in November [08]. Its scope includes AI, but whether it says anything about incidents will only be known once the signed text is published.
Among the documents read here, Singapore's are the only ones that single out the near miss, not just the incident, as what current rules miss. That gives it a head start on a question any regional framework will eventually have to answer.
Open questions
- QWhat is a near miss for an AI system? A refused agent action, a jailbreak caught by a filter, an evaluation failure before release? The pilot's first step is agreeing on categories, and the choice decides what the data can show.
- QWho reports: the developer, the organisation using the system, or the person who saw it? The Prime Minister put users at the centre. The EU AI Act places its reporting duties on the organisations that build or run a system.
- QCan voluntary sharing work alongside "legitimate security and commercial sensitivities"? A near miss is often embarrassing and rarely required. Without a duty to report, what makes it worth sharing?
- QWill human-approval logs become incident data? Every refused action in a government agent is a near miss already on record. Counting them is an open design choice.
- QDoes a pilot of willing countries produce data anyone can compare? Small cross-regional groups are easier to start. They are also the hardest to generalise from.
Every quotation above is from the source it cites. Sources 01 to 06 are the original documents. 07 and 08 are the best available and are marked.
- PM Lawrence Wong at the Virtual Leaders' Meeting on Frontier AI Safety co-hosted by Norway and Finland, Prime Minister's Office, 28 October 2026.Primary
- MDDI's response to PQ on reporting of serious and near-miss incidents involving autonomous AI systems, Parliament sitting of 4 November 2026.Primary
- Opening address by SMS Tan Kiat How at GovTech STACK Conference 2026, Day 2, GovTech, 5 November 2026.Primary
- MDDI's response to PQ on establishing Singapore as global leader in agentic AI security and assurance, 4 November 2026.Primary
- MDDI's response to PQ on assessment of risk of AI systems replicating across internet-connected devices, 4 November 2026.Primary
- Regulation (EU) 2024/1689 (the AI Act), Article 3(49), definition of "serious incident".Primary
- Artificial Intelligence Governance Bill to be tabled in Parliament early next year: Gobind Singh, Bernama, 29 October 2026.State news agency report
- Statement of the Chairperson of the ASEAN SEOM on the conclusion of ASEAN DEFA negotiations, ASEAN Secretariat, 2026.Primary · pre-signing