We build verifiable, adversarial RL environments for infrastructure and security operations which are tasks where an LLM agent has to actually fix a broken system or actually stop an intrusion, not just produce an answer that sounds right.
Most agent benchmarks grade responses with another LLM. We don't. Every environment we release resolves to a real, mechanically-checked outcome: the app either comes back healthy after the agent's fix, or it doesn't; the intrusion is either contained before it reaches its objective, or it isn't. No LLM judges, no rubrics, just a verifier that runs the actual system state.
In progress:
Both share one adversarial harness: a Red agent and a Blue agent, trained with GRPO on small open-weight models, competing against each other on real infrastructure rather than an abstract simulated network.
This grew out of SRE-Zero, our original RL environment for SRE incident-response tasks, built during the Meta PyTorch OpenEnv Hackathon.
Reinforcement learning environments have become one of the ways frontier labs actually train agentic models and most of the existing adversarial cyber-defense research (DARPA/TTCP's CAGE Challenges, CyberBattleSim, and similar) simulates an abstract network graph rather than a real, deployable stack. We're building against real infrastructure instead, because that's the gap between an agent that scores well on a benchmark and one that's actually useful when something breaks at 3am.
Early stage. Environments are in active development, see individual repos for what's currently working, benchmarked, and released versus still in progress. We open-source the environments and training code as they're ready; we don't publish results we haven't verified ourselves first.
Interested in a custom environment, a training/eval collaboration, or just want to see a demo reach out via indiumailabs@gmail.com or devaanshpathak.com.