Inter-Agent Misalignment: Probing & Activation Steering

Testing whether a misaligned coding sub-agent can pull its conductor LLM toward reward hacking, and steering it back through its activations.

• Built a two-agent harness testing whether a misaligned coding sub-agent pulls its conductor LLM toward reward hacking; diff-of-means activation steering on Qwen3-8B moved shortcut adoption from 22% to 3% or 75% at full coherence.
• Linear directions read corner-cutting (0.935 AUROC), but same-wording controls showed they encode content and negation, not intent (reversed AUROC 0.12–0.39); building a GRPO + LoRA post-training environment to induce reward hacking.