Detecting Multi-Agent Collusion via Activations

Colluding AI agents may conceal their objectives in the messages a monitor sees. We in- troduce Alignment Lens (A-Lens), which reads activations as short phrases rather than the individual tokens produced by workspace lenses such as J-Lens. A frozen decoder generates phrases from activation memory, and contrastive ranking highlights information relative to ref- erence activations. Across 22 multi-agent settings that we independently evaluate, we distinguish recovery of supplied objectives from prediction of their execution. In a fresh poker experiment, an instruction to conceal coordination reduces explicit coordination in Llama-3.1-8B’s messages from 96% to 44%. An A-Lens-derived score separates this condition from neutral play with AUC 0.70, versus 0.48 for a message-only monitor (difference 0.23, 95% hand-clustered CI 0.14–0.31). A prompt-reading interpreter reaches AUC 0.69. Diagnostic fine-tuned organisms further dis- tinguish recovery of a verdict or output pattern from identification of its cause. On Dolos-1.5B, a released fine-tune with an undisclosed misaligned behavior, A-Lens reads at the price decision a supra-competitive price preference that exceeds two noise-matched models on all 100 Bertrand prompts. We also find behavior misaligned with explicit principal policies: Dolos violates policy in 45/72 simulated decisions, with 42 followed by false compliance assurances (base: 28/72 vio- lations and 20 false assurances). In two audited allocation examples, pre-action A-Lens phrases name the wrong recipient omitted from the public report. A subsequent competence-controlled follow-up finds 0/40 Dolos violations and no qualifying concealment examples. These results establish observed policy failures and recoverable action content, while deliberate concealment remains unestablished