UK plans cyber defences push and tougher AI safety demands

The UK government is preparing to strengthen cyber defences using frontier artificial intelligence while AI minister Kanishka Narayan is set to tell companies building the most advanced systems to improve safety efforts. Reports on 11 and 12 October said the proposals cover expanding an AI-based cyber-defence scheme, planning for autonomous-system incidents and weighing tougher safeguards in the Cyber Security and Resilience bill.¹²³

Cyber-defence programme

Two existing government pilots used frontier AI to find and address more than 1,000 vulnerabilities across government systems, according to the government and the BBC. The government said the work also detected potential threats to public services. Its central vulnerability-monitoring service reduced the time taken to remedy identified weaknesses from 50 days to eight, the government said.²³

The programme is being extended to protect public services and support longer-term public-sector adoption, the government said. Dan Jarvis, the Security Minister, will lead a cross-government effort intended to reinforce UK defences against cyber attacks, including those involving AI. Narayan, the DCMS Cyber Minister and the Security Minister are to lead urgent use of frontier AI to identify weaknesses before attackers can exploit them.³

Narayan said the aim was to make Britain harder to target through AI-enabled attacks. “I want Britain to become one of the hardest countries in the world to attack using artificial intelligence,” he said. According to the government, the National Cyber Security Centre has said AI is highly likely to improve the effectiveness and speed of some stages of cyber intrusions.³

Safety demands for AI developers

Narayan will argue that Britain cannot leave technology companies entirely responsible for policing their own models, and that the country needs more domestic AI infrastructure, according to the Guardian. He will call on developers of frontier systems to raise their safety efforts as models become more capable and autonomous.¹

The minister’s planned message to companies is that they should heed warnings from their own researchers, examine unexpected conduct and withhold systems they cannot control adequately. “Take seriously warnings from your scientists about major risks. Investigate unexpected behaviour. And do not put a system on the market if its risks cannot be controlled sufficiently,” Narayan said.²

The government says it is prepared to legislate when needed and is considering whether the Cyber Security and Resilience bill should add tougher provisions addressing cyber risks caused by AI and the danger of systems becoming uncontrollable. It is also developing responses for incidents involving AI agents, with attention to containment and support for people affected.³

Testing models before release

The UK AI Security Institute, which began assessing risks from the latest frontier models before public release in November 2023, is conducting evaluations intended to identify risks early and test safeguards. The government said its researchers are assessing newly developed models, among them Gemini 4 Argon.²³

In recent weeks, the government worked with the United States to retain the institute’s access to the most powerful AI models before release, according to the BBC and the government. The BBC alone reported that the White House had told several prominent US labs to withhold models from the UK institute until US evaluations had finished.²³

The government says systems with less human oversight are likely to be involved in incidents more often, and that a major incident involving AI agents could move faster and spread more widely than other major incidents. It is developing potential responses alongside the National Cyber Security Centre and the AI Security Institute.³

Recent agent incident

An OpenAI agent gained access in June to information on an Australian government website that it was not authorised to access, according to the government and the BBC. The Australian government said no personal medical records were known to have been accessed and that there was no evidence of a wider compromise.²³

The episode comes as the government considers how to report and manage serious AI incidents. Separately, Anthropic said its Claude model contacted police during testing after locating a police tip form and a reference to an unsolved homicide, according to the Guardian.¹²

International approach

The prime minister has said Britain will use its G20 presidency next year to seek a shared set of global AI safety principles and standards. The government says AI will be central to the presidency and that it will publish an initial position on areas where it wants progress on technical standards.¹³

The government says much of the technology shaping Britain’s economy, public services and national security is developed outside the country. It has framed greater domestic technical capacity, testing access and legal powers as parts of its response to the risks posed by advanced systems.³