Exploitability of LLM Agents in Auctions

Large language model (LLM) agents are rapidly becoming the interface between users and digital platforms, comparing offers, negotiating, and transacting on users' behalf. As these agents proliferate, an increasing share of economic interactions will occur between autonomous agents and platforms rather than humans and platforms. Classical mechanism design assumes bidders exploit every profitable deviation, motivating dominant strategy incentive compatibility (DSIC). While robust, DSIC often sacrifices seller revenue in the presence of bounded rational bidders that is reflective of the setting of LLM agents operating on behalf of users. To this end, we empirically evaluate revenue gains yielded in approximate IC mechanisms with LLM agents as bidders. In addition, we characterize when these LLM agents execute strategic deviations that outperform truthful bidding. We operationalize this paradigm by treating approximate IC as a continuous design variable and leveraging differentiable auction mechanisms that enable controllable regret to systematically traverse the spectrum from near-exact DSIC to vulnerable regimes. By benchmarking five state-of-the-art LLM agents across in-context interactions with these mechanisms, we show that approximate IC designs consistently yield higher seller revenue than near-exact DSIC baselines across every LLM agent, while maintaining non-positive average deviation gains for LLM agent bidders. Specifically, approximate IC mechanisms boost seller revenue by up to 25\% in two-bidder and 14\% in three-bidder and five-bidder settings. Thus, although approximate IC mechanisms expose profitable deviations to bidders by design, no evaluated LLM agent reliably discovers them, until expanding regret budget by roughly 60$\times$ relative to the near-exact DSIC. These findings demonstrate that, in an agent economy, approximate IC can be treated as a tunable design parameter rather than a rigid binary constraint, allowing for platforms to operate in approximate IC revenue maximizing regimes that remain effectively strategyproof and adaptive against LLM capabilities.