BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.devconf.info//devconf-cz-2026//talk//EHUL8D
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-devconf-cz-2026-EHUL8D@pretalx.devconf.info
DTSTART;TZID=CET:20260619T153000
DTEND;TZID=CET:20260619T160500
DESCRIPTION:Securing a large language model today resembles an endless game
  of cat and mouse. Programmers try to manually write filters and prohibiti
 ons\, but all it takes is one creatively written prompt and the model obed
 iently generates dangerous content. Traditional defenses are inflexible\, 
 slow\, and attackers are always one step ahead. \nThis talk shows how to b
 reak out of this vicious circle. We introduce our open-source framework\, 
 which is used for systematic red teaming and testing models against 25 typ
 es of prompt-based attacks. We show how to analyze AI behavior under fire.
  On this basis\, we then introduce a new defense method based on genetic p
 rogramming. Instead of manually patching holes\, this "digital evolution" 
 automatically searches for optimal rules that strengthen the model and cre
 ate a defensive layer. All this without having to change a single paramete
 r in the model's weights. You will find out why evolutionary search for sy
 stem rules is more effective than an army of experts.
DTSTAMP:20260430T130835Z
LOCATION:A112 (capacity 64)
SUMMARY:Survival of the Safest: Automating LLM Defense with Genetic Program
 ming - Petr Kaška
URL:https://pretalx.devconf.info/devconf-cz-2026/talk/EHUL8D/
END:VEVENT
END:VCALENDAR