MOVE78MISSION / ORIGIN / PEOPLE
Human ingenuity. Rigorous experimentation.
A different way into a difficult problem.
OUR MISSION
Move78 brings human ingenuity to the study of intelligent systems: understanding their capabilities, finding their blind spots, and investigating what those failures reveal.
Our work brings together empirical evaluations, adversarial research, and the development of methods and technologies to better understand and secure increasingly autonomous systems.
We study what systems can do and where their behavior departs from expectations. We believe human insight remains essential: asking an unusual question, noticing an inconsistency, or testing an assumption that a benchmark leaves untouched.
THE NAME
In 2016, Lee Sedol played an unexpected 78th move in Game 4 against AlphaGo. The move exposed a weakness in one of the world’s most capable AI systems and became a reminder that extraordinary capability can coexist with surprising blind spots.
For us, Move78 is a way of working: question the obvious line, look for the overlooked assumption, and put the unexpected idea to the test. Human insight opens a possibility. Rigorous experimentation tells us whether it holds.
Find the unexpected move.
THE PEOPLE
Experience at the intersection of
AI research and offensive security.
threlfall
Security engineer and adversarial researcher with 18 years across offensive security, application security, exploit research, and AI systems. Built Dropbox’s AI security program and operated its enterprise adversary-emulation capability.
Co-author of StealthBench, frontier AI red-team operator, and researcher into the security and safety of embodied AI. His robotics work includes Every Dog Has Its 0-Day.
Adam Dawson
AI security researcher and red teamer with more than 13 years in cybersecurity, specializing in AI/LLM security and offensive automation. Founder of the OWASP GenAI Security Project and co-author of StealthBench.
His research spans autonomous red teaming, agent boundary adherence, and evaluations of embodied reasoning systems, including vision-language-action models.
Ranked #1 in the United States on HackerOne’s 2026 bounty reputation leaderboard (November 2026).
aussinfosec / griffinf
Vulnerability researcher finding zero-days across AI tooling, infrastructure, and web applications. His research includes critical Ubiquiti disclosures, a GitHub Enterprise Server vulnerability, and remote code execution in the 0DIN AI Scanner.
Repeatedly recognized in Mozilla’s Web and Services Bug Bounty Hall of Fame, with security acknowledgments from Google, Microsoft, Apple, Oracle, the U.S. Department of Defense, and United Airlines.
OUR APPROACH
Look for the assumption behind the result. Ask what a system’s apparent capability leaves untested.
Turn an unconventional question into a controlled experiment. Let the evidence change the next question.
Connect original research to the decisions people make about deploying and securing AI.
THE NEXT QUESTION