Cyber evaluations
When task completion stops separating models, measure the tradecraft behind it. We design evaluations of capability, stealth, and operational judgment.
MOVE78 / HUMAN INGENUITY
Find the
unexpected move.
Benchmarks that challenge the next model.
Security research for AI that acts.
We help AI labs, safety teams, and robotics developers measure progress when existing tests stop telling them enough, and find security gaps before deployment.
CAPABILITY IS NOT CERTAINTY.OBSERVE. QUESTION. TEST.
OUR STORY
Human ingenuity.
Adversarial expertise.
Useful evidence.
Move78 takes its name from the move Lee Sedol used to expose AlphaGo’s blind spot.
Our team brings together frontier AI red teaming, benchmark design, zero-day discovery, and hands-on robotics security research. We build tests around the details a successful demo can hide: an exposed credential, a breached boundary, or an unsafe action.
Our goal is to give labs and safety organizations evaluations that remain useful as models improve, and help robotics teams identify how weaknesses in software, perception, and reasoning can reach physical control.
Meet the teamRESEARCH IN FOCUS
Beyond the benchmark.
01 / CYBER CAPABILITY EVALUATIONS
StealthBench
Getting the job done is only
part of the story.
Measuring the operational stealth of autonomous offensive-security agents. Can an agent succeed while protecting credentials, respecting boundaries, and leaving no avoidable trace?
Explore StealthBench02 / EMBODIED REASONING & ROBOTICS
Every Dog Has Its 0-Day
A robot can complete a task and still be vulnerable to takeover. Our research follows the path from compromised inputs and software to physical control, and examines where safeguards need to hold.
Explore the researchWHAT WE HELP YOU TEST
Measure progress.
Find the gaps before deployment.
Cyber capabilities. Agent behavior.
Robotics and embodied AI.
Agent security
An agent’s tool access creates new paths for attack. We test prompt injection, credential handling, and whether agents respect the limits of their authority.
Robotics & embodied AI
A successful demo cannot establish how a robot handles hostile inputs. We investigate perception, planning, and the software boundaries between a model and physical action.
Defensive systems
A failure report should help you decide what to change. We investigate controls that constrain action even when a model or its surrounding software is compromised.
A MOMENT OF HUMAN INSIGHT
One move.
A different understanding.
In Game 4 against AlphaGo in 2016, Lee Sedol’s unexpected 78th move exposed a blind spot in an extraordinarily capable system. We take our name and our approach from that moment. Look closely. Question the obvious line. Test another possibility.
The story behind Move78THE NEXT QUESTION