Adversarial training infrastructure forAI × Cyber.
Offensive agents train on a clean signal: discovered flag, info leak, exploit executed, and the model learns with no human oversight. Defense has no equivalent signal, so the models that will run our networks are learning to attack far faster than they are learning to defend.
SentryStack builds the environments, platforms, tools, and verifiers that help AI perform its best on defense. Each one is a wargame: a real AI enabled attack chain, designed by cyber operators who built ranges for offensive testing, replayed against a defender on reproducible infrastructure making defensive outcomes machine-checkable, and trainable at scale.
Working on defensive RL, training AI, or want to know more?
A privileged view of the episode: attacker progress, mission uptime, collateral impact and remaining footholds, scored by the range itself rather than by a judge model.
A real attack chain replayed against a defender. The same compromised state is reset and branched, so isolating a host and terminating a process can be compared under identical conditions.
Executable worlds built by operators: hosts, identity, cloud, processes, services, networks and users, carrying a mission that stays measurable while it is under attack.
Built by operators





