OpenAI Supports Transparency. It Just Will Not Say When.

AI incident transparency was the one thing Singapore’s government asked the world’s leading AI labs for during Singapore International Cyber Week: a number.

On 13 October, Senior Minister K. Shanmugam told reporters he had met “all the major firms: OpenAI, Anthropic, Google, Amazon and others” at the event, and had put the same request to each of them. “Can we have full transparency immediately?” he asked. He wanted to know about AI incidents “at almost real-time, in fact, in real-time,” and said the companies “need to notify us early on when something happens.” He gave no figure. No hours, no days, no stated window. What he said instead was that “information needs to be given to us faster,” and that “the companies have expressed an attitude of cooperation.”

 

The Answer Every PMO Has Heard Before

That response is familiar to anyone who has sat on the other side of a steering committee table. A sponsor asks for a number: a date, a percentage, a ceiling. The delivery team agrees with the principle, praises the ask as reasonable, commits to cooperation, and somehow never attaches the actual figure. Singapore received that answer from three of the most capitalised companies on earth, in public, on the same afternoon.

OpenAI’s chief global affairs officer, Chris Lehane, gave the plainest version of it. OpenAI backs the principle of incident transparency, he said, and will alert affected Singapore agencies once it is confident an incident has occurred. What it would not do, on the same day, was attach a number to that promise. Lehane said OpenAI was “waiting to see if the Singapore Government is going to put an actual timeline on that.” He wants a general standard set through legislation, ideally aligned across governments over time, rather than a commitment OpenAI makes unilaterally to one jurisdiction. He also owned the company’s own record candidly: OpenAI found a security incident affecting Australian systems in August and did not tell Australian authorities until early September. Weeks, by the company’s own account, not hours.

 

AI Incident Transparency: A Different Route to the Same Non-Commitment

Anthropic took a different path to roughly the same place. Rather than negotiate a number, it signed something. On 13 October, Anthropic co-founder Ben Mann and Chng Kai Fong, permanent secretary at Singapore’s Ministry of Digital Development and Information, signed a memorandum of understanding, witnessed by Minister Josephine Teo. The agreement covers technical exchanges on model evaluation and testing standards with Singapore’s AI Safety Institute, research sharing through Anthropic’s Economic Index, and support for small business AI adoption. It names “incident response protocols” as a subject for exchange. It does not name a reporting deadline. By its own description, the MOU is a statement of intent, not a binding contract. Anthropic also opened its fifth Asia-Pacific office in Singapore that same week, a genuine commitment, just not the one Shanmugam had asked for across the table.

Two different responses, the same underlying move. OpenAI deferred the number to a future piece of legislation it hopes someone else writes. Anthropic deferred it into a framework of technical exchanges that has not yet produced one. Both are defensible positions for a company to hold. Neither is the answer Shanmugam asked for, and both arrived the same week he stood in front of reporters and asked for it directly.

 

The Question Has to Be Asked Twice

Every PMO has watched this play out at a smaller scale, a vendor who agrees enthusiastically with a governance principle while treating the actual commitment date as the one detail still under negotiation. The same substitution of language for commitment runs through how executives and finance teams talk past each other when they frame the same AI programme in different terms, where agreement on the words stands in for agreement on the figure. The lesson from Singapore is not that governments should stop asking. It is that the question needs to be asked twice, once for the principle and a second time, separately, for the number, because agreement on the first buys the other side time on the second.

Look back at the last vendor commitment your own organisation accepted without a date attached. Name the deadline it was meant to replace, and ask whether anyone ever went back and asked for it.

Scroll to Top