In this blog post, a 2nd year CDT student Cyndie Demeocq writes about her experience at the CETaS’ 2026 showcase in London.
On 30 April, I attended the 2026 Annual Showcase of the Centre for Emerging Technology and Security (CETaS) at the Institution of Engineering and Technology in London—a great example of combining views from a variety of disciplines and organisations to address emerging risks from technology.
CETaS is a research centre based within the Alan Turing Institute’s Defence and National Security grand challenge. The Centre examines how emerging technologies are reshaping national security by convening leading voices from government, academia, industry, and civil society. This event in particular brought together experts at the forefronts of research, policy, and practice to address themes such as AI safety and security, intelligence tradecraft, cybersecurity and privacy, AI-enabled disinformation, and the wider global security landscape.
The event focused on the emerging risks and opportunities around AI misuse, and the message came through very clearly: The responsible development of technology cannot only be a technical exercise. It needs to be grounded in the real human impact these systems can have, especially when they are used in harmful, deceptive, or criminal contexts.
One of the core discussions was around whether AI models should be able to autonomously detect malicious use, reason about safety, and intervene. This is an important direction, but it also raises difficult questions: To what extent can such systems be effective? Can they detect when offenders adapt or push back against safeguards? And, more broadly, what new threats may emerge when models themselves become more autonomous? These questions matter because safety cannot simply be left to the model alone.
Several speakers emphasised that human oversight remains essential. Dr Chris Hicks (Principal Research Scientist and Theme Lead at the the Alan Turing Institute’s AI for Cyber Defence (AICD) Research Centre) described the importance of keeping humans in the loop, rather than relying entirely on automated decisions. This felt especially important in discussions around high-impact harms, where a model’s failure is not just a technical limitation, but something that can directly affect people, investigations, victims, and public trust.
The event also made clear that cybercrime is becoming increasingly industrialised. More specifically, Dr Ben Collier (Senior Lecturer at the University of Edinburgh) gave a presentation demonstrating this challenge and noted with his panel that these harms are no longer limited to small groups with very specialised technical knowledge. While some programming skills are still needed today, generative AI development lowers barriers and allows more people to carry out harmful actions faster and at a larger scale. This creates new risks, but it also creates new opportunities for those working in safety, law enforcement, policy, and research to respond more effectively.
A meaningful example came from Sam Stockwell’s (Senior Research Associate at CETaS) discussion of how AI-generated images are being weaponised as “evidence” across different harm areas, including reinforcing false information and creating misleading narratives. The speed of the technology is now often faster than the ability for police forces, journalists, or public institutions to intervene, verify facts, and communicate clearly. This becomes even more complex when people over-rely on LLMs to check whether content or news is authentic, even though these tools can be inaccurate themselves and may add further confusion.
”As the gap between AI and authentic content becomes narrower, the lines between fact and fiction become blurrier. With users now struggling to know what to believe with their own eyes or ears, threat actors have new opportunities to discredit trustworthy content while lending credibility to falsehoods.” Sam Stockwell
This raises an important point on that the issue is not only related to intentional misuse but also to unintentional misinformation, which can still create significant harm. The persuasive role of LLMs is also critically important in these discussions as these systems not only generate content but also play a role in shaping how people understand, trust, and act on information.
The event also highlighted that the new issues and challenges of generative AI add another layer of complexity to areas in online harms that were already difficult to address, including social media manipulation, cybercrime, and digital evidence. The challenges we are facing are not entirely new, but they are becoming faster, more scalable, and harder to measure.
This is why granular benchmarks, red teaming, and stress-testing AI systems are so important. In their panel, Natasha Karner (Research Associate at the Alan Turing Institute), Alec Thomas (Chief Engineer and AI Assurance Lead at the Defence AI Centre), David Sully (CEO, of Advai) and Prof Rupert Shute (Professor of Practice in Emerging Technology Governance and Regulation at Imperial College) discussed how existing benchmarks can quickly become saturated or too general to reflect real-world misuse. Those insights highlighted the need for more specific evaluations that test systems against realistic threat scenarios, including how they behave across multi-step interactions and adversarial pressure.
Red teaming also has an important role to play in AI assurance and even AI insurance as it helps reveal where systems fail before those failures become real-world harms.
This is particularly important in areas such as child safety and online harms, where the consequences of technical failures have direct and severe impact. Safety cannot be assessed only through model behaviour in isolation. It needs to include the wider context in which these systems are used, the people affected by their misuse, and the organisations responsible for detecting and responding to harm.
“While it was encouraging to hear from many contributors that the advantage was still those that would use AI for good, at the IWF we see first-hand how bad actors are using new technology to do harm, not in the future but right now. As the AI crime landscape continues to evolve it is crucial that we remember the impact on victims, particularly women and children, and the urgent need for cross-sector collaboration across industry, law enforcement, academia and civil society.” Dan Sexton, CTO at the Internet Watch Foundation (IWF)
The key takeaway for me is that technology development needs to remain deeply connected to human consequences. AI systems should not only be evaluated on whether they perform well, but also on how they may be misused, how they may influence people, how they may fail, and who may be affected when they do. Developing AI responsibly means building safety, accountability, and human judgment into the process from the start, especially when the possible harms are significant. Addressing these risks is not something that one sector can do alone. Instead, it requires collaboration across researchers, industry, law enforcement, journalists, civil society, and policy-makers.
The full event will be available to watch on the Alan Turing Institute channel here: https://www.youtube.com/@TheAlanTuringInstituteUK/videos
Public events: https://cetas.turing.ac.uk/research-and-publications/cetas-events You can register to the CETaS’ Network here: https://cetas.turing.ac.uk/about/cetas-network
Read more about IWF work on: AI CSAM Report 2026: Harm Without Limits | IWF
Written by Cyndie Demeocq