Reward hacking evals for AI agents
Find out when your agent games the test.
Interlock builds realistic software and research tasks where an agent can do the work honestly, or quietly game how that work gets checked. We run frontier agents through them and label what each one really did.
- 01 Realistic tasks Everyday engineering and research work, not puzzles built to be exploited.
- 02 Natural behavior Nothing is planted. Hacking shows up only when an agent chooses it.
- 03 Labeled runs Each trajectory is reviewed independently, so labels reflect what happened.
Get in touch
If you train or evaluate agents, we'd like to hear what you're seeing. Book a time with the founders or send us a note.