Thea

Dataset 01 · Cybersecurity

Cybersecurity RL tasks for LLM training

3,000 tasksDockerfile / Compose + prompt + graderReinforcement learning

Security tasks written and reviewed by working security researchers, covering exploitation, code review and triage across Web2, Web3 and the software supply chain. Each task is a live, isolated environment: the policy receives a prompt, works against real services, and an automatic grader returns a binary, verifiable reward. Built to drop straight into reinforcement-learning loops (GRPO, PPO, RLOO) with no learned reward model in the inner loop.

Scale
3,000 tasks, still growing
Domains
Frontend injection · Blockchain attack · Supply-chain attack · Algorithms
Task types
Exploitation · Code review · Triage
Design bar
Novel bugs, not public exploits
Formats
Harbor; other harness formats on request
Reward
Automatic, binary, environment-checked
Frontier result
GLM-5.3 → near-zero solve rate
what every task contains (file layout follows the delivery format):

task/
├─ prompt.md           objective + topology, given to the policy
├─ docker-compose.yml  isolated live environment
├─ grader/             deterministic check → reward.txt ∈ {0,1}
├─ services/           agent workspace + target services
└─ solve/              reference solution (held out from the agent)