Policy Brief: Artificial Intelligence and Cybersecurity
Also available in PDF
Artificial Intelligence is, surprisingly, not a new field - the term was first coined at the Dartmouth Summer Research Project on Artificial Intelligence in 1956. Only recently, however, has there been sufficient computing power and available data, alongside technical progress, such as transformers, to create systems that exhibit something resembling intelligence. With the release of the first publicly available large language model, ChatGPT, on 30 November 2022, the dynamics of AI reached a new level, and no end is in sight. It is hard to conceive what comes next, given that the technology appears to leap from milestone to milestone at extraordinary speed.
This rapid technological development, coupled with massive investment - predominantly from the US and China - is creating challenges on several fronts. Traditionally, states respond to new developments with regulation. Given the current pace of change, however, regulating specific technologies seems futile, as such regulation is perpetually overtaken by events. Experience shows that new capabilities typically become broadly available within six to nine months of their introduction. Rather than attempting to regulate specific technologies and constantly falling behind, societies need to develop the means to manage rapid change. This means evaluating established values and norms in light of new technologies and addressing the underlying issues directly.
The same applies to the technical and security communities, which must also adapt to rapid change.
This policy brief identifies some key concerns raised by recent advancements in AI.
There is more than frontier models
Public opinion is largely shaped by the capabilities of leading-edge frontier models and by hopes of containing them. Experience shows, however, that within a short time, smaller models exhibit comparable capabilities. For example, after Mythos was announced, it took only a few weeks for researchers to reproduce some of its security-relevant results by orchestrating older models1.
Recommendation: Treat proprietary frontier models as a preview of technology that will soon be broadly available. Don’t base policy decisions on short term developments.
AI will disproportionately affect those below the security poverty line2
The latest models are not only extremely capable at finding vulnerabilities, but also at creating exploits and potentially deploying them. While large organisations will face pressure, they are likely to cope, supported by AI tools and adequate budgets. This is not the case for organisations living below the security poverty line, those without the resources to field skilled security teams or deploy specialised technologies. This includes the majority of small and medium-sized enterprises (SMEs), including those operating critical infrastructure such as power utilities and hospitals. Such organisations face significantly elevated risk of being targeted by state-sponsored actors or criminal organisations.
Recommendation: Prioritise solutions and policy that protect vulnerable organisations as soon as possible. Such organisations can no longer afford to rely on being under the radar.
AI will hit critical underfunded software
Today, up to 96%3 of all software shipped contains open source components. Yet very few of these projects are well funded. The same is true for much of the “invisible” software, driving much of our critical infrastructure. At the same time, AI models are generating quality bug reports at rates 50 to 100 times higher than before. Open source projects, however, are left with the same number of volunteers fixing bugs in their spare time. This has long been a concern; AI now amplifies it exponentially.
Recommendation: Develop measures to help protect critical open source components embedded in software supply chains. We encourage community driven approaches to address these issues.
AI affects a broad set of actors
Past regulation and policy focused primarily on commercial entities. The dynamics of the security space, however, are driven by a much broader set of actors: open source authors, security researchers, bug bounty programmes, civil society, end users, and the operators of foundational internet services such as DNS, BGP, and NTP.
Recommendation: Consider all actors relevant to global cyber stability.
AI empowers
AI is not only a threat — it also empowers actors who previously lacked resources. It is essential that access to AI technologies is available to a broad range of actors, not only wealthy organisations. Open weight and smaller models have an important role to play here. Meaningful access also means access to knowledge and expertise, not merely training courses, but the creation of environments in which ideas and capabilities can flourish.
States should engage in long-term, meaningful capacity building in line with the UN Framework on Responsible State Behaviour in Cyberspace.
