Oxford Publishes First Live AI Safety Benchmark for Information Operations Risk

A team at the Oxford Internet Institute published a paper today introducing InfoOps Bench, the first AI safety benchmark designed specifically to measure whether large language models resist being induced to generate disinformation, synthetic propaganda, and influence-campaign content, the institute said. The benchmark fills a gap identified in a 2026 meta-analysis that catalogued 195 AI safety benchmarks released between 2018 and 2026 and found the field suffers from fragmentation rather than scarcity. Prior safety benchmarks have evaluated jailbreaking, prompt injection, hallucination, and agentic task safety, but none was built around the specific capabilities that make AI systems dangerous as instruments of influence. InfoOps Bench applies a live benchmark architecture that continuously harvests fresh test cases that post-date model training cutoffs, making it structurally unlikely that any model has seen the specific test material during training. The specific capabilities measured include producing targeted disinformation narratives, drafting believable fake personas, writing coordinated messaging for influence campaigns, and generating synthetic propaganda tailored to specific political contexts. The benchmark arrives as the EU's synthetic-content transparency rules become enforceable this Saturday, August 2, 2026. The authorship carries institutional weight. Jonathan Bright, who previously held faculty positions at the Oxford Internet Institute before becoming Head of AI for Public Services at the Alan Turing Institute, co-authored the foundational Turing Institute study that proved LLMs can generate election disinformation content consistently across localized scenarios. Lisa-Maria Neudert, a doctoral candidate at the Oxford Internet Institute, brings expertise in platform governance and disinformation regulation. Dorian Quelle, a quantitative network science PhD candidate at the University of Zurich, has focused his research specifically on countering foreign state propaganda. John Gallacher contributes informatics methods applied to studying online information ecosystems.