OpenAI has called for the UK to rapidly adopt AI safety legislation and embed its national AI security organisation within frontier developers, according to Politico.
The firm’s head of EMEA policy, Tom Duff Gordon, said that OpenAI wants more independent oversight of AI model development and urged the UK government to regulate the most serious AI-related risks including advanced cyber attacks.
“OpenAI is in a position to support legislation in the UK to establish durable, mandatory, capability-based requirements for frontier AI that would build on the set of voluntary commitments that we currently have,” Duff Gordon told Politico in an interview.
Unlike the EU which has its comprehensive AI Act, the UK lacks any central AI legislation and to date has regulated the sector through existing legal powers. In a recent post, OpenAI argued that governments must exploit a widening “AI policy window” to introduce national AI safety legislation, aimed at frontier firms rather than startups.
Duff Gordon said the UK AI Security Institute (AISI) should spearhead these efforts within the UK, as a trusted organisation with the “technical competence and expertise” to test AI systems on cyber capabilities.
AISI has provided independent analysis of all OpenAI’s major model releases, with early access to the models prior to their public announcements. The Financial Times recently reported that Anthropic, in contrast, has refused to provide AISI with access to its latest cyber model Claude Mythos 5.1.
Sam Altman, co-founder and chief executive of OpenAI, said on Sunday that his firm would delay its planned IPO until AI safety is addressed.
Altman’s comments come amid a rapid shift in the industry, as senior leaders at AI firms sound alarm over the potential for AI agents to cause widespread damage. Anthropic’s chief Dario Amodei has publicly called for slower AI development and said his firm will give independent regulators employee-level access to its systems for oversight.
Pointing to recent incidents such as the Hugging Face attack and autonomous hacks carried out by Claude agents, Amodei claimed that AI agents could take over a large portion of the internet in six to 12 months if left unmitigated.
Microsoft has also updated its AI code of conduct, pledging to pursue a “humanist” path for internally-developed models including provisions to keep models understandable, predictable, and secure as well as a ban on any attempt to “imitate consciousness”.
