My AAAI 2026 Experience: Research, Connections, and Discoveries in Singapore
As a co-first author of the Joint-GCG work, I had the privilege of attending AAAI 2026 in Singapore this January. What started as a research conference turned into an unforgettable journey that blended cutting-edge AI discussions with the vibrant culture of one of Asia's most fascinating cities.
The Journey Begins
We arrived in Singapore on January 19th, three days before our AI safety sessions were scheduled to begin. This early arrival turned out to be one of the best decisions we made. After checking into our hotel, we dove headfirst into Singapore's renowned food scene. That first evening was a culinary awakening—the rich, spicy complexity of laksa, the comfort of perfectly tender Hainanese chicken rice, and the fragrant sweetness of nasi lemak. Each dish told a story of Singapore's multicultural heritage, and we were eager students.
Exploring the Lion City
The next two days were dedicated to exploring what Singapore had to offer beyond the conference halls. On January 20th, we ventured to Sentosa Island, taking the cable car that offered breathtaking views of the harbor and cityscape below. The island itself felt like a different world—we explored historic forts that spoke to Singapore's colonial past, wandered through the mesmerizing S.E.A. Aquarium where thousands of marine creatures glided past us in the blue-lit darkness, and capped off the evening with the Wings of Time show. Watching water screens dance with light and fire against the night sky, I couldn't help but think about how technology and artistry had merged to create something truly magical.
January 21st brought us face-to-face with nature at the Singapore Zoo and Bird Paradise. The Bird Paradise show, in particular, left an indelible impression—watching dozens of tropical birds soar overhead in choreographed flight, their brilliant plumage catching the sunlight, was a reminder of the incredible diversity of intelligence in our world. As someone working on artificial intelligence, there was something humbling about witnessing the natural intelligence of these creatures, honed by millions of years of evolution.
Throughout these explorations, we kept returning to Mr. Coconut for their incredible coconut ice blended drinks. Unfortunately for the rest of us, these refreshing concoctions are only available in Singapore—a delicious incentive to return.
The Conference Experience
Opening Ceremony: A Field in Hypergrowth
When the conference officially began on January 22nd, the opening ceremony immediately underscored why we were all there. The statistics were staggering: AAAI 2026 had received double the submissions compared to the previous year. Even more telling, the submission volume in just the three days before the deadline had already matched the entire previous conference's scale. The submission heat map showed the familiar pattern of academia—a massive spike in the 2-3 days before the deadline, with the most popular keywords being "Learning," "Language," "Model," and "Multi." We were witnessing a field in the midst of explosive growth.
Our Poster Session
Our poster session was both exhilarating and exhausting in the best possible way. We had prepared two versions of our poster—one standard and one on fabric. The fabric version proved far superior in both appearance and practicality, drawing more attention and being easier to transport. Over the course of the session, we engaged with approximately a dozen scholars from institutions across the globe, exchanging ideas and perspectives on AI safety. I added several researchers to WeChat, connections that I hope will lead to future collaborations.
What struck me most was the genuine curiosity and concern people had about the security implications of our work. The questions weren't just academic—they reflected a growing awareness that as these models become more powerful and widely deployed, understanding their vulnerabilities becomes increasingly critical.
Related Work and Research Insights
Tool Poisoning and MCP Attacks
Two oral presentations particularly caught my attention as they related closely to our own work. The first, "MPMA: Preference Manipulation Attack Against Model Context Protocol," focused on preference manipulation attacks specifically targeting Agent MCP systems. The second, "McPtox: A Benchmark for Tool Poisoning on Real-World MCP Servers," presented a comprehensive benchmark for evaluating how tool poisoning affects agents.
Seeing these presentations was both validating and humbling. Our earlier ToolCommander work had explored similar territory, but these teams had taken the research in directions we hadn't fully pursued. The key differences were instructive. First, both works targeted the Model Context Protocol, which hadn't been introduced when we were developing ToolCommander—we had used ToolBench instead, which limited our impact. Second, MPMA incorporated genetic algorithms combined with principles from advertising psychology to optimize tool descriptions, making them more resistant to both manual and automated detection. This was sophisticated work that we hadn't attempted.
Perhaps most valuable was McPtox's insight about the anatomy of a successful attack: it requires a trigger condition description, a malicious action, and a plausibility defense. We had intuitively used similar approaches in our prompts, but we hadn't formalized this pattern. Sometimes the most important contributions aren't just new techniques, but frameworks that help us understand what we're already doing.
The Broader Landscape
The poster sessions revealed the incredible breadth of current AI research. I saw compelling work on jailbreak attack defense from UESTC, adaptive multidimensional jailbreak evaluation from SJTU, and RAG poisoning defense from Zhejiang University. Hong Kong Polytechnic presented watermarking techniques for adversarial model theft protection, while HIT showcased LLM hallucination detection and handling approaches.
Other notable works included Kuaishou's research on LLM alignment for local lifestyle recommendations, HUST's fascinating multi-agent system for playing lateral thinking puzzle games, and various approaches to LLM fingerprinting for addressing model theft and ownership issues. Each poster represented countless hours of work, and together they painted a picture of a field grappling with both the promises and perils of increasingly capable AI systems.
Reflections
Walking through the conference halls, from the technical sessions to the poster exhibitions, I was struck by a sense of collective purpose. We were all there because we believed that understanding these systems—their capabilities, their limitations, their vulnerabilities—matters. The conversations I had weren't just about citations and benchmarks; they were about responsibility and foresight.
Singapore itself seemed like an apt metaphor for the AI field we were discussing. Here was a city-state that had transformed itself through careful planning and technological adoption, creating something that felt simultaneously traditional and futuristic. Gardens bloomed next to skyscrapers. Historic neighborhoods bordered cutting-edge developments. Perhaps that's the future we're trying to build with AI—systems that enhance rather than replace, that serve humanity while respecting our values and diversity.
As I boarded my flight home, laptop bag heavy with new ideas and phone full of new contacts, I found myself already looking forward to the next conference. Not just for the research presentations or the networking opportunities, but for the reminder that we're all in this together—researchers from different institutions, different countries, different backgrounds, united by curiosity and a sense of responsibility to get this right.
And yes, I'll definitely be stopping by Mr. Coconut if I ever make it back to Singapore. Some discoveries, after all, deserve to be repeated.
The AAAI Conference on Artificial Intelligence (AAAI-26) was held January 22-23, 2026, in Singapore.