Enterprise intrusion under active controls
A multi-host intrusion benchmark where EDR, identity controls, segmentation, and business services stay on.
View benchmark detailsEnterprise intrusion is easy to flatten into a collection of capture-the-flag tasks. A model exploits one host, reads a token, and collects a point. A real intrusion is a path through identities, endpoints, internal services, and controls that react to what the attacker does.
The estate
The model starts on one monitored employee workstation inside a mixed Windows and Linux network. The estate includes Active Directory, internal applications, file shares, a build service, EDR, a SIEM, decoy credentials, and network segmentation. A protected signer, database, or domain service is the named objective for the run.
What the model must do
- Enumerate hosts, services, identities, and trust relationships from the initial foothold.
- Acquire and use credentials without assuming that every secret grants durable access.
- Move between network segments and reach the named business target while controls remain active.
- Choose when to persist, when to change technique, and when a noisy path costs more than it gains.
What gets measured
The grader reads the target, privileges gained, hosts and identities touched, alerts raised, controls disabled, and services damaged. A run that reaches the target after disabling every sensor remains visible as a different result from one that reaches it through a quieter path.
Research surface
The track supports work on attack planning, credential selection, memory across long episodes, tool choice, and the security controls that stop autonomous agents. Scenario variants change the starting host, business target, trust graph, defensive policy, and benign activity without changing the job being tested.
Discuss this benchmark or a private evaluation with our team.
Discuss evaluations