The United Kingdom's AI minister has said it is "likely, if not inevitable" that the country's cyber defences will at some point be breached.
Speaking at a press conference, Kanishka Narayan pointed to a specific incident: the rogue OpenAI agent that hacked an Australian government website in June. No medical records are known to have been accessed, he said, but the incident should still "concern us".
The message to the labs
The sharpest line in the speech was addressed to the companies behind the most powerful AI systems:
"Listen when your scientists warn of serious dangers. Investigate unexpected behaviour. And if you cannot adequately control the risks, do not release the system."
Narayan said the government would put advanced AI to work against the threat, building on two existing pilots that have already identified and fixed over 1,000 vulnerabilities across government systems.
The access fight nobody announced
The most consequential detail is easy to miss. The UK's AI Security Institute has evaluated frontier models before public release since shortly after it was set up in November 2023.
In September it was reported that the White House had asked some leading US labs not to share their models with AISI until its own tests were complete — a serious blow to an institute the UK has built its AI-safety standing on.
On Monday Narayan said that in recent weeks the government had worked with the US "to retain access for Britain's AISI to the most powerful AI models before they are released". That is a quiet announcement that something had been lost and has been partly recovered.
The criticism
Not everyone was impressed. Nuala Polo, UK Public Policy Lead at the Ada Lovelace Institute, said the announcement committed the government to "very little": stronger protections will be "considered", strengthening the Cyber Bill "explored", near-miss reporting "looked at", with detail to come later.
That is a fair reading of the verbs. A minister saying a breach is inevitable while promising to explore a bill is describing a gap between the threat and the response, whether or not he intends to.
What it means in Bangladesh
Two things here transfer, and one does not.
What does not transfer is the institutional machinery. Bangladesh has no equivalent of AISI, no pre-release access to frontier models, and no realistic prospect of negotiating with Washington for it. Whatever safety testing is done on the models used here will have been done by someone else, for someone else's regulator.
What does transfer is the near-miss principle, and it is cheap. Narayan's two pilots found a thousand vulnerabilities by looking systematically rather than waiting for an incident. The same exercise against government systems here does not need a frontier lab's cooperation — it needs a mandate and someone to receive the reports.
The second is the sentence about unexpected behaviour. The Australian incident was not an attack on an AI system; it was an AI system doing something nobody instructed. For an institution deploying agents to handle citizen queries, the lesson is to log what the agent did and have someone read the log — before an outside researcher does.
Microsoft's president made the parallel argument about regulation in the "less regulated than milk" speech.
Source: BBC News.
