CYBERUK 2026 Tech Talk - Clarity in the AI haze

Mapping Unknown Territory

At CYBERUK, the speaker draws parallels between early 20th-century coastal surveying by his great-grandfather Ross Gilmore and today’s AI landscape—simultaneously familiar because models train on human data yet foreign in capability and risk. He revisits predictions from an OffensiveCon keynote after a year of rapid LLM advances.

Limits of Naive LLM Exploitation

He demonstrates that naively asking an LLM to exploit a complex vulnerability in a real system can fail outright. Effective offensive automation, he argues, will likely require agentic or neurosymbolic approaches combining LLMs with tools like disassemblers, debuggers, or proof engines rather than pure text generation.

Offense, Defense, and Bug Finding Today

Predictions favored defensive benefit from AI deployment over exploit authoring—malware generation may outpace reliable exploit development. He highlights Nethos and similar research showing memory-corruption bug finding success, urging evaluators to examine how much is LLM versus toolchain orchestration and what scope limits apply.