One Probe Is Not Enough: The Bipolar Geometry of Deception in LLM Negotiation Agents

AI control protocols require trusted monitors that can flag deceptive behavior from untrusted models, but current approaches rely on LLM-based monitors that only see text outputs and require a full additional inference call per output. We present an activation-based monitoring protocol that reads the model’s internal representations directly, detecting deception via a single linear probe a single dot product on activations already computed during the forward pass, requiring no additional model inference. Using a Concordia-based multi-agent negotiation framework, we deploy Gemma-7B-IT agents across three scenarios, capturing 28.5K residual stream activations via TransformerLens. The probe achieves AUC 0.81–0.82 at Layer 14, significantly exceeding all baselines. Three findings are directly relevant to control protocol design: (1) emergent and instructed deception share nearly identical internal directions (cosine similarity 0.76–0.93), meaning probes trained on cheap synthetic data detect dangerous spontaneous deception; (2) commission deception (lying) and omission deception (withholding) are geometrically anti-correlated (−0.76 to −0.80), creating a quantified blind spot that any single-probe monitor will miss; and (3) the deception subspace is 1-dimensional, removing one direction drops AUC from 0.79 to 0.09, meaning the simplest possible monitor works. These results provide a concrete, deployable component for AI control: a cheap, fast, internal state monitor with a known and addressable failure mode.