'Months, Not Years' — Why Five Governments Just Warned Your Board on AI
On 22 June the Five Eyes cyber agencies issued a rare joint statement. Their words: the timeline is not years, it is months. They didn't address it to IT teams — they addressed it to boards and executives. That warning landed alongside five more stories, and together they say one thing. Treating AI as the IT department's problem is over.
The Promise
- Delegation is genuinely working inside the labs. OpenAI published research showing Codex now generates more than 99% of the work its own staff produce with AI, and the share of users handing the system a full day's job went from 2% to 26%.
- Mistral shipped OCR-4, a document-reading model built to run inside your own infrastructure so sensitive files never leave the building. Regulated industries can get state-of-the-art document AI without sending a single page to someone else's cloud.
- Five intelligence services do not align on a public statement for something speculative. The most senior warning of its kind I've seen went to boards rather than to IT — which is the right address.
The Risk
- Anthropic named a major Chinese conglomerate in a model theft case: linked operators used roughly 25,000 fake accounts to run almost 29 million interactions with Claude models. The technique is distillation — train a cheaper model on a stronger one's outputs and copy the capability for a fraction of the cost.
- Your AI supply chain no longer ends at software and cloud regions. It now includes rented intelligence, and rented intelligence can be copied, stripped of its safety controls, and turned against you.
- Oracle's headcount fell by about 21,000 in a year, and its annual report filed 23 June states that AI has resulted in reductions to its workforce, with restructuring costs around $1.8bn. An interview can be walked back. A securities filing cannot.
- GPT-5.6 launched as a limited preview to about twenty trusted partners at the government's request, with broad release following a federal review of its cyber capability. The same gate that pulled Anthropic's models, and OpenAI rates all three tiers high for both cyber and biological risk under its own framework.
The warning, and who it was addressed to
On 22 June the Five Eyes cyber agencies issued a rare joint statement. Their words: the timeline is not years, it is months.
They did not address it to IT teams. They addressed it to boards and executives. Frontier AI, they said, is about to reshape both attack and defence, and cyber resilience is now a question of business continuity.
It is the most senior warning of its kind I have seen. Five intelligence services do not align on a public statement for something speculative.
Rented intelligence can be copied
Anthropic named a major Chinese conglomerate in a model theft case. Linked operators used roughly 25,000 fake accounts to run almost 29 million interactions with Claude models.
The technique is called distillation: you train a cheaper model on the outputs of a stronger one, copying the capability for a fraction of the cost.
Your AI supply chain no longer ends at software and cloud regions. It now includes rented intelligence — and rented intelligence can be copied, stripped of its safety controls, and turned against you.
From asking to delegating
OpenAI published research on how its own people use Codex. Inside OpenAI, that one tool now generates more than 99% of the work their staff produce with AI, and the share of users handing the system a full day’s job has gone from 2% to 26%.
People are moving from asking to delegating. When you delegate, your job changes. It becomes oversight — reviewing, supervising, and signing off on work you did not do.
A price on that shift, in a securities filing
Oracle filed its annual report on 23 June. Headcount fell by about 21,000 people in a year, and Oracle wrote that AI has resulted in reductions to its workforce. Restructuring cost around $1.8bn.
That sentence is the news. Oracle is one of the first companies of its size to put AI layoffs in writing to regulators. An interview can be walked back. A securities filing cannot.
The gate, and the countermove
OpenAI unveiled GPT-5.6 in three tiers, and you cannot use it yet. At the government’s request it launched only a limited preview to about twenty trusted partners, with broad release following a federal review of the model’s cyber capability. The same government gate that pulled Anthropic’s models — and OpenAI rates all three tiers high for both cyber and biological risk under its own framework.
While OpenAI sits behind that gate, Mistral shipped OCR-4, a document-reading model built to run inside your own infrastructure so sensitive files never leave the building. The pitch is sovereignty. As US export controls and access gates tighten, that is a sharpening commercial argument rather than only a political one.
The week, assembled
Five governments warning boards. A model theft case. Jobs becoming oversight. AI layoffs in a filing. A frontier model behind a government gate. And a sovereignty play.
This was the week AI stopped being the IT department’s problem.
I spent 25 years in cybersecurity watching the same arc. Boards delegated security, called it handled, then watched the breaches show up in proxy statements and director lawsuits. AI is running the same arc, faster.
The question for your next meeting is not whether you use AI. It’s who is accountable when the AI you rented, deployed, or delegated to gets it wrong.