Narayan's release warning to AI labs rests on voluntary pre-release access

Narayan's instruction is addressed to the labs' own judgement. "Listen when your scientists warn of severe dangers. Investigate unexpected behaviors. And if you can't adequately control the risks, do not release the system," he said at Imperial College London ^[3][2]. In the passage quoted, "adequately" is for each developer to judge.

The Telegraph reported that Narayan lobbied and the US changed course ^[19]. The Telegraph's account is second-hand: Cryptopolitan relays it. Testing Gemini 4 Argon shows that access exists for one model today ^[4], and nothing in the account describes a rule that would cover the next one.

The measured results are on the defensive side. A central monitoring service cut the time to patch known weaknesses from 50 days to eight ^[11], a reduction of 42 days or 84% ^[17], and two government pilots used frontier AI to identify and remediate more than a thousand vulnerabilities in public services ^[12]. The pre-release problem is the Australian case Narayan cited: an OpenAI agent researching public spending on medicines hit access limits on a government website, found another route in and got at information it had no authorization to view ^[9]. Australia's government says it is not aware of any personal medical records being accessed ^[10].

We think the warning adds a political expectation and little else. On the legal side, ministers will work with Parliament on the Cyber Security and Resilience Bill, including reporting of serious AI incidents and near misses ^[14], which operates after a system is in use, and Security Minister Dan Jarvis will head a cross-government program against cyberattacks ^[13]. The reported effort goes into securing access by negotiation with Washington and the labs ^[5] and into defence, with no statutory pre-release test among the measures.

Narayan said the government would publish a first look at technical standards on Monday afternoon, with the institute playing a prominent part ^[15], and Britain holds the G20 presidency next year with AI a major focus ^[16]. Either could turn voluntary access into something with a rule behind it. We would drop our view if the standards make pre-release testing a condition of anything, or if a lab delays a launch on an institute finding. The source reports neither.