Skip to main content
Talk

Artificial Intelligence and Security in Large Language Models at the Colombia World Security Summit

Dr. Omar Costilla Reyes spoke on AI and security in large language models at the Colombia World Security Summit in Medellín on May 17, 2023. The in-person talk was part of the summit’s inaugural edition.

Poster for the academic agenda of the Colombia World Security Summit, held in Medellín on May 17–18, 2023. It names ESU and Plaza Mayor Medellín as organizers. Credit: Image courtesy of Dr. Omar Costilla Reyes.
Academic agenda for May 17, 2023, at the Colombia World Security Summit, showing the talk and the expert panel that followed. Credit: Image courtesy of Dr. Omar Costilla Reyes. View image source.

An in-person talk at the inaugural summit

The first Colombia World Security Summit took place on May 17–18, 2023, in the Gran Salón of the Plaza Mayor convention center in Medellín. The City of Medellín announced it as an event organized with Plaza Mayor and set within the 40th anniversary of ESU, the city’s security and urban-solutions company. It combined a trade exhibition with an academic program for businesses, local and regional governments, startups, and engineering and technology students, with dozens of invited experts, themes that ranged from cybersecurity and data analytics to artificial intelligence, and free admission on registration.

Dr. Omar Costilla Reyes announced in Spanish that he would speak in person on the future of security and artificial intelligence. The summit’s agenda graphic lists him as an MIT research scientist for a TED-style conference, “Presente y futuro de la inteligencia artificial y seguridad” (the present and future of artificial intelligence and security), on May 17. Afterward he posted about the talk, on the future of artificial intelligence and security in large language models, and noted that a full video would be published.

Why language-model security matters

Large language models are neural networks trained on very large collections of text so that they can predict, generate, and summarize language, and many are now built into assistants that draft messages, answer questions, and write software. Security discussion about them generally covers two linked concerns. One is protecting the model and the systems around it: crafted inputs can make an assistant ignore its instructions, sensitive information can leak through prompts or training material, and answers can be fluent but wrong. The other is misuse, for example the large-scale production of convincing phishing messages or false content. Dr. Costilla Reyes describes the discussion as asking how such tools can be made useful while guarding against misuse.

Context and related work

On the same day’s agenda, the talk was followed by an expert panel on technological migration and the future of security in Latin America, with other speakers. Related pages on this site include his listing as an invited presenter on MIT initiatives in Latin America and artificial intelligence at the Organization of American States’ Tenth Regular Meeting of the Inter-American Committee on Science and Technology (May 2023), and the NeurSym and MIT CSAIL team’s research on language models trained on programs, which he shared in May 2023.

Topics