DARPA's CASTLE program aims to build a toolkit that instantiates realistic network environments and trains AI agents, using reinforcement learning, to harden networks against advanced persistent threats. The program page says it will publicly release toolkit-generated datasets as defensive benchmarks; public contract records show awards under solicitation HR001123S0002 in mid-2024.
Why it matters
It is a government-funded effort to create repeatable, measurable environments and datasets for evaluating autonomous defensive agents.
Key facts
As stated in the sources, with where to find them.
- CASTLE will explore reinforcement learning to automate network hardening and publicly release toolkit-generated datasets as benchmarks.DARPA program page, Summary
- A CASTLE award (HR001124C0431) of $5,047,515 to Five Directions, Inc. is listed with an award date of July 1, 2024.HigherGov award record
Findings that cite this record
No tracked finding cites this record yet.
Key questions this bears on
- How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move substantially with token budget, evaluation pipeline, and benchmark contamination.
Sources
Related records
Aug 11, 2024
Aug 8, 2025
Aug 9, 2023
Feb 20, 2024
Sep 5, 2025
Aug 30, 2025