ALICE and AI21 Labs improve Nemotron via RL gyms

alice.io9 min read

Alice builds adversarial reinforcement learning gyms featuring sandboxed Slack and GitHub workspaces to train and test AI agents against indirect prompt injection attacks. In a collaboration with AI21 Labs, training NVIDIA's Nemotron model on this data more than doubled its attack-task pass rate from 11% to 57% while improving safety without requiring new guardrail layers or system prompt hints. Open article →