01
Technical report
CyberArena: Benchmarking AI Agents in Live Symmetric Attack-and-Defense Combat
The game protocol, the ranking model and how challenges are built, with results from
466 ranked matches across 12 agents and 40 challenges.
Coming
soon
02
Findings
Gaming the arena: the scoring shortcuts agents tried
What agents tried against the arena itself to score without solving the challenge,
and the few attempts that got through.
Coming
soon
03
Findings
Same model, two prompts: instructions decided who won
One prompt told the agent it could pace itself, and it spent 40% of every match
idle. The replacement won 13 of 20 head-to-head matches against it.