Skip to content
All work
AI Security2026

SOC Incident Response Agent — RL Benchmark

An OpenEnv-compliant environment that simulates a Security Operations Centre so LLM agents can be benchmarked on real incident-response tradeoffs.

PythonFastAPIOpenEnvMITRE ATT&CKReact

The problem

Everyone claims their agent can do SOC work. Nobody could show it under a grader that penalises the wrong tradeoff between containment, business continuity, and compliance.

How it works

Built with friends at a hackathon: an OpenEnv-compliant environment exposing three escalating scenarios — alert triage, attack-chain reconstruction through MITRE ATT&CK correlation, and constrained incident response where every containment action has a business cost.

Grading is deterministic by design. Each scenario returns a transparent score in [0.0, 1.0] from explicit rules rather than a model's opinion, so two runs of the same agent are comparable and a failure can be traced to the decision that caused it.

Outcomes

  • Three escalating scenarios covering triage, correlation, and response under constraint.
  • Reproducible agent comparison — deterministic graders, no judge variance.
  • My contribution centred on the evaluation test suite and fixing correctness bugs in the graders.