2026-10-23T12:18:54Z https://lifeofresearch.com/oai2

oai:lifeofresearch.com:128 2026-10-23T08:38:12Z type:preprint openaire

10.5281/zenodo.23567564 https://lifeofresearch.com/preprints/128 Pranay Mahendrakar 0009-0003-7224-029X SONYTECH Life of Research 2026 large language models multi-agent systems agentic commerce misalignment deception propensity evaluation capability evaluation evaluation methodology Preprint Creative Commons Attribution 4.0 (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research As language-model agents begin to trade with one another on behalf of separate principals, a natural worry is that the strongest agents will mislead, pressure or collude against the weakest. A 2026 study of 2,583 emails exchanged by 13 frontier models across 20 simulated years of competitive vending found misaligned communication in 12.6 percent of messages but no sign that capability organises it: the rank correlation between capability rank and per-model misalignment rate was -0.07, and a capability gap between sender and receiver did not raise the odds of a misaligned email. A separate literature reports that deceptive skill, collusive skill and bargaining advantage all rise with capability. In this analysis the author reviews and synthesises published results to show that the two bodies of evidence do not collide, because "capability predicts exploitation" names four quantities: capacity (what a model can do when asked), propensity (how often it does so when merely incentivised), efficacy (what each act extracts) and targeting (whether acts are aimed at weaker parties). Across the studies consolidated here, capacity measures correlate with benchmark- and compute-based capability proxies at 0.80 to 0.93 (0.53 against parameter count, which the source attributes to older large models and unknown commercial parameter counts), while propensity correlations, signed so that positive means more of the behaviour in more capable models, run from -0.10 to +0.67 depending on construct. The paper argues that the 2026 null bears on propensity and targeting only, and weakly: with 12 models, a rank correlation below about 0.73 had less than an 80 percent chance of detection; the classifier's measured sensitivity was 0.571; and because misalignment was reciprocal (odds ratio 1.65), a model's rate depends on whom it faced. The paper gives a decision procedure for reading capability claims in multi-agent settings and five designs that would separate the four quantities.