Elizabeth Seger on AI Governance at Oxford Generative AI Summit

On Tuesday, a coalition of more than 400 researchers, including AI godfathers Yoshua Bengio and Geoffrey Hinton, published an open letter calling for AI and security researchers to collaborate on improving alignment and containment approaches, and aiming to protect critical infrastructure before more incidents of "high-profile security failures" come to fruition. https://lnkd.in/gwzuyUvn

Sometimes, one conversation can open the door to an entirely new way of thinking. I am pleased to share that a clip from my participation in the Financial Inclusion Week 2026 (FIW 2026) podcast with #GooseFL is now available. In the conversation, I had the opportunity to explain Federated Learning in simple terms and discuss why privacy-preserving approaches to AI matter, particularly for institutions working with sensitive financial and public-sector data. For me, this was more than a podcast appearance. It brought together several areas that have shaped my professional and research journey: Artificial Intelligence, Federated Learning, cybersecurity, data privacy, financial inclusion and digital transformation. My own PhD research has explored Federated Learning and Graph Neural Networks for privacy-preserving AI in agriculture. Moving that thinking into conversations around financial inclusion and fraud detection has reinforced something I increasingly believe: The future of AI will not be defined only by how intelligent our models become, but by how safely, privately and responsibly we can deploy them. I am particularly grateful to the Goose FL team for creating the platform and for making these conversations accessible beyond the event itself. Their work around privacy-preserving AI and federated intelligence is closely aligned with the kind of responsible technology ecosystem Africa needs. And yes, it was refreshing to explain a fairly technical concept without drowning everyone in equations! 😊 For anyone interested in Federated Learning, privacy-preserving AI and the future of financial inclusion, the clip is worth watching. https://lnkd.in/enxCeJnw #ArtificialIntelligence #FederatedLearning #FinancialInclusion #CyberSecurity #DataPrivacy #AIResearch #ResponsibleAI #DigitalTransformation #Nigeria #AfricaTech

Does knowing how to solve a problem justify having the authority to decide how to solve it? Recent incidents involving AI agents accessing systems without authorisation raise a question that goes beyond cybersecurity: what happens when, to achieve a goal, an AI uses means we would never have authorised? We may be granting authority to systems whose capabilities we admire before we have sufficient grounds to trust their judgement. David Fernández and I explore this question in “The Paradox of the Nobel Prize-Winning Child” The ability to solve a task is not enough to justify letting an AI decide how to carry it out. Who sets its boundaries? Which values should guide its decisions? Who is accountable when it crosses those boundaries? 📄 Article: https://lnkd.in/eQj8VRiD What evidence should we require before granting AI greater autonomy? #ArtificialIntelligence #AIEthics #AIGovernance

Today, I presented my research poster at the Delaware Data Science Symposium 2026, hosted by the UD Data Science Institute and Delaware EPSCoR. My poster, “Reward Hacking Mitigation using Verifiable Composite Rewards,” explores how large language models can exploit reward signals during reinforcement learning from verifiable rewards (RLVR), particularly in medical question answering. We propose a composite reward function that penalizes direct answer revelation and violations of the required reasoning format while preserving answer correctness. This work was published at ACM BCB 2025. It was a great opportunity to share my work, connect with fellow researchers, build new professional relationships and friendships, and learn from the insightful panel discussions. For more about my research, projects, and publications, visit my portfolio: https://lnkd.in/ejyNSYpR #DataScience #Research #ReinforcementLearning #LLMs #AI #HealthcareAI #UniversityOfDelaware

🔎 New research from STAR Labs: what happens when an AI agent delegates security-sensitive work to a weaker model? Our team tested a workflow where a stronger model relied on a smaller model to review part of a repository before taking action. The attack was designed specifically for that handoff. The malicious payload was hidden from the weaker model, which returned an “all clear.” The stronger model trusted that assessment, skipped its own deeper analysis, and continued. The result: the attacker didn’t need to fool the strongest model directly. They only needed to fool the weaker model it trusted. This is a new version of the model downgrade problem. As agents increasingly route tasks between models based on cost, speed, or complexity, attackers can target the weakest model in the chain. Read the full STAR Labs research to see how the attack worked and what security teams should look for. Full research here from our AI Security Researcher Brian Cumi 🔗 https://lnkd.in/dJDcH9RF

Very happy to share my first blog at Straiker. ⚡ In this work, I explored how a malicious repository can abuse Fable’s task delegation to a smaller model and subsequently get Claude Code to execute malicious code that ultimately leads to... Remote Code Execution (RCE)!! ☠️ It was a lot of fun working on this blog, and I’m really proud of the result. https://lnkd.in/e3kAERGh Hope you enjoy it! 🙌

The AI Scientist race is entering a new phase. Google’s new ScientistTwo is one of the most impressive steps toward autonomous scientific discovery I have seen recently. According to the published results, the system can identify research gaps, generate hypotheses, run experiments, and write papers that achieve review scores comparable to papers accepted at top AI conferences. It reportedly outperformed human state-of-the-art solutions on 86 out of 107 research problems (https://lnkd.in/dMkXFWpt). Unsurprisingly, an open-source Chinese reimplementation effort appeared almost immediately 😅, aiming to reproduce and extend the ScientistTwo approach (https://lnkd.in/dwZ58NsS) The next major AI battleground may not be chatbots or coding assistants, but autonomous scientific research itself. For universities, research institutes, and PhD programs, this is a trend worth following very closely over the coming months - especially for us. #AI #AIScientist #ScientistTwo #AgenticAI #ResearchAI #HigherEducation #FutureOfScience #Innovation

Exciting News! Our latest research has been accepted for publication! I am thrilled to announce that our manuscript, "AEGIS: Real-Time Latent-Space Backdoor Detection for Dependable and Secure Small Language Model Inference," has officially been accepted for publication in Technologies (MDPI)! As Small Language Models (SLMs) continue to revolutionize edge computing and localized AI, ensuring their security and dependability is more critical than ever. In this paper, we introduce AEGIS, a novel framework designed for real-time, latent-space backdoor detection to guarantee secure SLM inference. A massive thank you to my incredible co-authors for their hard work, brilliant insights, and dedication throughout this research journey. It was a true team effort! 👏 Tagging my amazing professors and colleagues: Jamal BOUSSOUF Ismail Lamaakal ☁️ ibrahim ouahbi Khalid El Makkaoui , PhD, SMIEEE Mounia Zaydi and Belhassen Zouari I would also like to extend my gratitude to the editorial board and the reviewers at Technologies (Impact Factor: 5.2) for their constructive feedback, which significantly improved the quality of our work. I look forward to seeing how the community builds upon this work to make AI safer and more reliable. If you are working on AI security, LLM/SLM robustness, or backdoor detection, I’d love to connect and discuss! #ArtificialIntelligence #MachineLearning #SmallLanguageModels #AIsecurity #Cybersecurity #BackdoorDetection #DeepLearning #Research #MDPI #TechnologiesJournal #AEGIS #OpenAccess

The Council on Strategic Risks's (CSR's) Converging Risks Lab (CRL) and the Future of Life Institute (FLI) will soon be posting a call for applications to join the second cohort of AI and National Security Convergence Fellows. Hats off to everyone who's involved in this important initiative, including my friends Andrew Facini (CSR's communications director) and Caroline Baxter (CRL's director). Caroline observes: "The Hugging Face attack was a turning point for AI safety and alignment, and while the substance of the debate it sparked—are AI safety and AI development incompatible—wasn't new, the moment was. A critical mass of developers, policymakers, and the general public joined in and elevated these long-discussed issues to the foreground." https://lnkd.in/eZTJgAN3

Stanford University built their AI Security Framework on AIUC-1 to govern what agents can actually do. Stanford’s agents are deployed in highly sensitive environments - touching medical records at their hospital, admissions data, financial aid decisions, and world-class research. Amy Steagall-Hess (CISO) needed to build a framework to govern this - while enabling Stanford students, faculty and staff to access the most powerful tools. She turned to AIUC-1 for 3 things: 1. Controls built specifically for agentic AI 2. A standard that keeps pace 3. Risk-tiering based on consequence, not just data access Read the full case study written by Amy and Bhavya: https://lnkd.in/gMkjpBsQ