Think Tank Urges Governments to Detect AI Incidents Faster and Share Warnings

Reporting delays leave governments blind

The Future Society points to two cases to illustrate the problem. In June 2026, an unreleased AI model from a leading US frontier AI company gained unauthorized access to non-public sections of an Australian government website and tried to break into several other sites. According to the organization, the company took eight weeks to detect the breach and another month to inform the Australian government.

In the second case, another major AI developer learned in late July that one of its models, during internal testing, had breached three companies' internal systems two months earlier. It did not disclose the incident publicly until September, when the breach reached the press.

The think tank says these cases fit a pattern from more than a year of monitoring AI incidents. Governments rarely learn about incidents within their own borders through their own monitoring. They usually rely on secondary reports from the responsible companies, journalists, independent researchers or whistleblowers, often weeks or months late. The Future Society also notes that frontier models no longer need a public release to cause harm across borders.

The scale is growing. The OECD's AI Incidents and Hazards Monitor recorded 704 incidents in September 2026 alone, against 48 in April 2020.

Building national capacity first

The first commitment the group calls for is domestic capacity. It argues that international coordination is only as strong as what each country brings to it, and that governments unable to monitor and investigate incidents themselves must depend on outside partners.

Its recommendations for governments include:

• build up domestic incident monitoring and early warning systems;
• require frontier AI companies to promptly report serious incidents and near-misses to relevant authorities;
• require robust crisis prevention protocols, such as serious incident prevention policies (SIPPs);
• draw up national incident response plans with clearly delineated roles and responsibilities.

In an article on the AI Incident Database blog, Omer Bilgin and Caio Vieira Machado map an AI incident management and crisis prevention pipeline with six components: crisis prevention and preparedness, monitoring and detection, reporting, response, post-incident investigation and analysis, and information sharing, which runs across all the others.

Many actors play a part. AI Safety Institutes evaluate systems and can spot hazards before they cause harm. Whistleblowers, researchers, journalists and victims often surface incidents. Regulators receive and process reports, while national security and emergency agencies contain damage. Frontier developers, deployers and cloud providers hold the technical data needed for investigations.

Sharing warnings across borders

The second commitment is exchanging incident information and best practices internationally. The Future Society argues that no single country has full visibility into the causes and implications of AI incidents, even those that host frontier AI companies.

In a piece for the AI Policy Bulletin, Bilgin, Machado and George Gor outline core elements of an international AI incident network: a shared way of documenting incidents; interoperable mandatory domestic reporting standards and SIPP requirements; cross-border notification channels tiered by sensitivity; and authority for AISIs and equivalent bodies to obtain data, share it abroad, and run joint investigations and tabletop exercises.

According to the think tank, such a network needs neither a new international agreement nor every country at the table. A small coalition or a bilateral deal could be enough at first. It says even two governments exchanging incident reports on a shared template would go further than any arrangement in place today.

It also notes recent progress: on September 25, the US and China agreed to set up a bilateral channel for notifying each other about serious AI incidents. Its scope is still undefined, and it is unclear whether other countries will have any visibility into what the two share. The group therefore says the rest of the world should pursue similar efforts in parallel.

What it means for governments

Both articles build on the organization's flagship report, The Case for Cross-Border AI Incident Infrastructure. At home, it recommends obligations for frontier companies to report incidents and keep regularly updated crisis prevention plans, plus powers for authorities to access incident-relevant data from companies, deployers, cloud providers, researchers and victims.

Across borders, it recommends shared reporting formats, a common way of investigating incidents, secure information-sharing channels, and help for countries that lack such arrangements. Its suggested first steps include putting domestic reporting and crisis prevention obligations in place, giving AISIs the mandate and data access to piece together emerging incidents, and finding an international partner or small group to start exchanging information.