Between the two of us we have spent most of a career on the offensive side of cyber: one as an Army cyber warfare officer, one building offensive tooling with NSA and U.S. Cyber Command. What keeps us up now is not that models are learning to attack. It is the rate.
Offense has a clean training signal and defense does not. That asymmetry is not a detail. It decides which capability compounds.
Why offense is pulling ahead
An offensive agent gets told to take the box. It runs, and a script checks the outcome: the flag dropped or it did not, the exploit executed or it did not, the data left or it did not. The reward is +1 or 0, exact and verifiable, and the loop runs unsupervised at whatever scale you can afford.
Now score a defender. It patched a service, killed a process, changed a firewall rule. Did it work? You need to know whether the attacker was actually stopped, which part of the chain broke, whether persistence survived, and whether the system the defender was protecting still does its job. Today a person answers that, once per run. So defensive training runs at human speed while offensive training runs at machine speed.
What we are building
SentryStack builds the environments, platforms, tools and verifiers that give defense a reward a machine can check. Each environment is a wargame:
- Executable worlds. Hosts, identity, cloud, processes, services, networks and users, carrying a mission workload that stays measurable while under attack.
- A real attack chain. Authored by operators who built ranges for offensive testing, not assembled out of puzzle boxes.
- Branching rollouts. The same compromised state, reset and replayed, so competing responses are compared under identical conditions.
- A privileged verifier. Attacker progress, mission uptime, collateral impact and remaining footholds, scored by the range rather than by a judge model.
What comes out is a defensive reward with the property that made offensive training work: objective, reproducible, and nobody has to watch it happen.
Why now
- The compute has already moved. Frontier labs have shifted the majority of their training budget from pre-training to reinforcement learning, and demand for environments is outrunning the supply of people who can author them.
- Cyber defense has no leader. RL environment companies in other domains are already worth billions. The defensive cyber slice is open.
- Training domains are not interchangeable. An environment that teaches a model to find bugs does not teach it to hold ground under a live adversary.
Real-world defense takes people who have been on offense. We recruit from the community we came from, and we are building the infrastructure we wish had existed when we were on the other side of it.
Attack stopped. Mission intact.