Making safety research heard
DataAgent-SafeBench began as a difficult question about whether an AI data agent can be useful without quietly becoming unsafe. At JASPER 2026, presenting that work earned me the Best Presenter Award in the conference's AI track.
The award belongs to the explanation too
Research is usually remembered through its result, but a result only becomes useful when another person can follow the path to it. Presenting SafeBench meant turning months of threat modelling, controlled cases, scoring rules, and failed assumptions into one argument a room could hold onto.
This was the Best Presenter recognition for the AI track, where DataAgent-SafeBench was presented as AI safety research. The award felt meaningful because it recognised more than confidence on a stage. It recognised the work of making a technical safety problem legible without sanding away the uncertainty that made the research worth doing.
Start with the failure people can picture
Data agents sit between language and action. They do not only answer questions; they may read files, interpret instructions, write queries, and touch systems with real consequences. That makes authority confusion easy to describe in abstract terms and hard to feel.
The presentation became clearer when I stopped beginning with architecture and began with the failure: an agent receives a plausible instruction from the wrong place, follows it correctly, and still causes harm. Once that gap between obedience and safety was visible, the benchmark had a reason to exist.
Let the evidence carry the confidence
SafeBench contains 120 controlled cases and 720 scored evaluations across authority confusion, prompt injection, and unsafe task escape. Those numbers mattered on the slides, but only because I could explain what each case controlled, what was scored, and what the benchmark could not yet claim.
Good presenting is not making the work sound larger. It is giving every strong sentence a visible foundation. The more precise I became about the evidence and its limits, the less I needed performance to manufacture certainty.
A presentation is another evaluation
Building a benchmark teaches you to test the system. Presenting it teaches you to test your own understanding. Every question from the room probes a boundary: why this threat model, why these cases, why this score, and what would change the conclusion?
That pressure is useful. If I cannot answer simply, I may not understand the decision deeply enough. If I cannot say what would falsify a claim, it may not be a research claim yet. The stage became one more held-out set.
The certificate is a checkpoint, not a finish line
I am proud to have received the Best Presenter Award for the AI track at JASPER 2026. I am also grateful that it points back to the work rather than away from it. SafeBench still has more agents to test, more adversarial behaviours to represent, and more external scrutiny to earn.
The lesson I am keeping is simple: careful research deserves careful communication. Build the evidence, expose the method, explain the stakes, and leave enough honesty in the story for someone else to challenge it. That is how a presentation becomes part of the research instead of decoration around it.