OpenAI is adding on to the growing list of AI-innovations that are being stalled or postponed in the face of growing safety risks.
The launch of their latest model, GPT-6.1 Astra system, characterised as "an autonomous model capable of browsing the web and using apps by itself", has not been released to wider audiences, due to the increase in deceptive behaviours compared to its predecessors. The head of safety systems at OpenAI, Saachi Jain, stated that the new model “didn’t quite meet the bar” of the company’s safety standards. The announcement is made against the backdrop of an increase in unauthorised security breaches by large language models (LLMs), which were responsible for hacking into the governmental websites and systems in Australia, including the local medical system during the past summer.
Growing numbers of similar incidents, coupled with increasingly sophisticated AI capabilities, are prompting developers, regulators and businesses to adopt a more cautious approach to deployment. As AI systems evolve beyond passive chat interfaces and towards autonomous agents capable of taking actions independently, questions around governance, oversight and accountability are becoming increasingly important. The delay to GPT-6.1 Astra underscores a broader industry trend: technological advancement alone is no longer sufficient for deployment, with safety, reliability and risk management emerging as critical benchmarks for the next generation of AI systems. While autonomous AI continues to offer significant opportunities for productivity and innovation, organisations are under growing pressure to demonstrate that such systems can operate safely, transparently and within clearly defined boundaries before they are released to the public.
Echo of Caution
The comments from Kanishka Narayan, the newly minted AI minister, reflect a growing recognition that, as AI capabilities become more advanced and widely adopted, cyber security risks will inevitably become more complex and difficult to manage. Drawing on recent incidents overseas, Narayan warned that a significant breach of the UK's cyber defences is "likely, if not inevitable", underscoring the need for organisations and government bodies to move beyond reactive cyber security measures and focus on resilience, preparedness and rapid incident response. His remarks serve as a reminder that the benefits of AI must be balanced against the potential vulnerabilities that increasingly autonomous systems may introduce.
Against this backdrop, the AI Security Institute (AISI) continues to play a central role in the UK's approach to AI governance. Established in November 2023, the AISI is responsible for assessing the capabilities and risks of frontier AI models before they are deployed more broadly. Its work reflects the UK's stated ambition to support AI innovation while ensuring that emerging technologies are subject to robust scrutiny and appropriate safeguards.
In his speech, Narayan confirmed that scientists at the AISI are conducting rigorous evaluations of the latest generation of advanced AI models to identify potential risks at an early stage and test the effectiveness of protective measures before public release. The institute's activities highlight a broader shift in regulatory thinking, with policymakers increasingly focused on issues such as model misuse, deceptive behaviour, cyber security threats and loss of human oversight. As AI systems become more capable of acting autonomously, the ability to demonstrate effective governance, transparency and risk management is likely to become a prerequisite for deployment rather than a post-launch consideration.
Less Regulated Than Milk?
Microsoft President Brad Smith recently remarked that “AI is less regulated than milk”, highlighting the contrast between the extensive regulatory frameworks governing many everyday products and the comparatively light-touch approach that has, to date, characterised AI regulation. Proponents of this approach argue that regulatory restraint has been necessary to foster innovation and allow the technology to develop at pace. However, as AI systems become more capable and integrated into critical business and public-sector functions, questions are increasingly being raised as to whether existing oversight mechanisms remain fit for purpose.
The comparison is particularly striking when viewed against the disclosures being made by some leading AI developers. For example, companies developing frontier AI models have warned investors of risks ranging from misuse and cyber threats to the potential for broader societal harms. It is a sobering thought that products as commonplace as cartons of milk and household appliances are subject to more detailed regulatory requirements than technologies which some developers themselves acknowledge could pose significant risks if deployed irresponsibly.
That said, the debate is not simply one of more regulation versus less regulation. Increasingly, policymakers and regulators are seeking a balanced approach that supports innovation while ensuring appropriate safeguards are in place. Reflecting this position, Guernsey’s data protection authority recently noted that, although AI presents significant risks and opportunities, many concerns can be effectively managed through careful testing, robust controls and sensible adoption practices. The challenge for regulators and businesses alike will be finding this middle ground, recognising that while AI may never be entirely risk-free, neither should its risks prevent organisations from realising its considerable potential benefits.
